{
  "schema_version": 1,
  "date": "2026-11-19",
  "week_start": "2026-11-13",
  "edition": "Kickoff dossier",
  "title": "Which AI changes should change a lesson?",
  "summary": "The first Academy AI Review traces new model releases, access on everyday devices, and the gap between a leaderboard score and a task a learner can verify.",
  "coverage": "The reporting window is November 13–19, 2026. Earlier releases establish a July–November baseline, with an older evaluation method included for context. The selection favors changes that affect access, prompting, task choice, cost, or verification. Source dates and the evidence-check date are recorded separately. The benchmark numbers come from the named evaluators. The Academy exercises below are proposed tests, and their answer keys are supplied for independent scoring.",
  "protocol": "For each comparison, save the fixture, exact prompt, date, model ID or visible model name, product, enabled tools, effort and number of attempts. When available, log elapsed time and actual charges. Keep unknown costs as unknown. Score against the answer key before checking model identity. For a small pilot, run the same fixture three times in fresh sessions per setup, report the individual scores and failures, and use another fixture before generalizing. A learner-directed verification step provides separate evidence of the learner’s skill.",
  "leads": [
    {
      "text": "For beginners, two practical changes arrive on devices they may already own. ChatGPT shows dictated words before submission, and eligible Android Go phones can run Gemini. A short task with an answer key can work across both interfaces.",
      "sources": [
        "chatgpt",
        "gemini-go",
        "gemini-mobile"
      ]
    },
    {
      "text": "For model selection, GPT-6.2 and Gemini 4 Neon make a good case for recording the exact task and effort setting. Independent results changed with the task, and the price comparison changed with the rate used.",
      "sources": [
        "aa-openai",
        "aa-gemini"
      ]
    },
    {
      "text": "For developers, two prompt migrations deserve different treatment. OpenAI changed where lasting instructions belong in a GPT-6.2 Responses request. Anthropic documents how Fable 5.2 handles reasoning effort and short output budgets.",
      "sources": [
        "openai-migration",
        "claude-migration"
      ]
    },
    {
      "text": "For local AI, a published 1.1 GB model file offers a concrete first experiment. A different project’s 180B-on-Mac demonstration is an interesting follow-up with a very different memory requirement.",
      "sources": [
        "liquid-files",
        "stream-demo"
      ]
    }
  ],
  "findings": [
    {
      "id": "gpt62-surfaces",
      "section": "Models and prompting",
      "provider": "OpenAI",
      "date": "2026-11-17",
      "evidence": "Documented release",
      "availability": "API release and desktop preview in Work and Codex on eligible plans. Managed seats depend on workspace settings.",
      "title": "Where learners can actually select GPT-6.2",
      "change": "OpenAI released GPT-6.2 Astra and Sol through the API and previewed both in ChatGPT Work and Codex desktop. The ordinary Chat picker kept its existing choices at launch. In the API both models take text or images and return text. OpenAI positions Astra for its more demanding work and Sol as the lower-cost option.",
      "implication": "A learner can follow the right steps on the wrong screen and never see the model. A learner who uploads a diagram can ask for an explanation, but an image-editing exercise also needs to identify the tool that produces the edited image.",
      "try_this": "Put the product, plan and selected model beside a diagram-reading example. Give the learner the diagram legend and ask them to verify each label in the answer.",
      "sources": [
        "openai-release",
        "openai-api",
        "chatgpt"
      ]
    },
    {
      "id": "gpt62-prompt-migration",
      "section": "Models and prompting",
      "provider": "OpenAI",
      "date": "2026-11-17",
      "evidence": "Documented behavior",
      "availability": "Advanced track for API users. The behavior concerns the request sent by an application.",
      "title": "Reusable instructions have a new home in the Responses API",
      "change": "For GPT-6.2, OpenAI’s guide says top-level instructions are appended after the input. Its advice for reusable instructions is to put them in an initial developer message so they are eligible for prompt caching. Editing earlier messages can also cause the API to discard prior reasoning items.",
      "implication": "A prompt template can still appear to work while a repeated task gets slower or consumes more tokens. A beginner can learn to state an audience and deliverable. The developer guide should also show where a program supplies those instructions.",
      "try_this": "Try a stable instruction for explaining electronics to a beginner, followed by two different questions. Compare the second request’s usage with a run that edits the earlier instruction. Record the API model and request layout.",
      "sources": [
        "openai-migration",
        "openai-release"
      ]
    },
    {
      "id": "fable52-access",
      "section": "Models and prompting",
      "provider": "Anthropic",
      "date": "2026-11-09",
      "evidence": "Documented release",
      "availability": "Fable is offered on the Claude Platform and supported cloud platforms. Mythos requires organizational verification.",
      "title": "One underlying Claude model has two access paths",
      "change": "Anthropic describes Fable 5.2 and Mythos 5.2 as the same underlying model with different safeguards and access conditions. Fable’s standard API rates at launch were $8 per million input tokens and $40 per million output tokens. Eligible paid app seats have their own usage rules.",
      "implication": "A family member using the Claude app and a developer using an API account have different choices and billing. An Academy comparison can use the accessible Fable path and plainly identify eligibility when discussing the counterpart.",
      "try_this": "Give learners the same two-page public project brief and ask them to find three contradictions. Record which product and model they could access and score each contradiction against the brief.",
      "sources": [
        "claude-release"
      ]
    },
    {
      "id": "fable52-effort",
      "section": "Models and prompting",
      "provider": "Anthropic",
      "date": "2026-11-09",
      "evidence": "Documented behavior",
      "availability": "Applies to Fable 5.2 API migration. Effort controls exposed by other products can differ.",
      "title": "A shorter answer and a smaller reasoning budget are different requests",
      "change": "Anthropic’s migration guide reports that prose asking for less deliberation did not reliably reduce thinking at medium effort or higher. The supported effort setting controls that budget. A very low output-token cap can be spent before the model returns any answer text.",
      "implication": "A three-bullet answer is an output requirement. The amount of computation the service spends is a separate setting. Conflating them can make an otherwise sensible beginner prompt look broken when it is transferred into an app.",
      "try_this": "Keep the same request for three inventory labels. On an eligible interface with effort control, vary effort alone, then record the final labels, elapsed time and any missing output.",
      "sources": [
        "claude-migration",
        "claude-release"
      ]
    },
    {
      "id": "haiku55-routing",
      "section": "Models and prompting",
      "provider": "Anthropic",
      "date": "2026-10-07",
      "evidence": "Provider positioning",
      "availability": "Available as claude-haiku-5-5 on the Claude Platform and through supported cloud providers.",
      "title": "A smaller Claude model gives a routing experiment a starting point",
      "change": "Anthropic introduced Haiku 5.5 for quick, high-volume work such as summarization. The release also changed Sonnet 5.5 cache-read pricing. A guide that quotes a single family price or uses one time estimate for every task can become inaccurate while model names stay familiar.",
      "implication": "The practical question is when a smaller model meets a written success condition. Classifying short service notes can be scored cheaply. Resolving a contradiction across manuals deserves a different test.",
      "try_this": "Compare two available models on ten invented service notes with known labels. Record missed labels and retries before considering a larger model. Treat any routing rule as a hypothesis until the test is repeated.",
      "sources": [
        "claude-small"
      ]
    },
    {
      "id": "gemini4-neon",
      "section": "Models and prompting",
      "provider": "Google",
      "date": "2026-11-18",
      "evidence": "Documented release",
      "availability": "Paid Gemini API tier. The Gemini app and Antigravity rollout includes eligible Google AI Pro and Ultra subscribers.",
      "title": "Gemini 4 Neon adds a middle choice for complex work",
      "change": "Google’s launch lists four thinking levels, a one-million-token context window, and launch API prices of $3 per million input tokens and $15 per million output tokens, before cache discounts. The model accepts mixed materials such as documents and charts for analysis.",
      "implication": "A large input allowance describes what can be supplied. Whether a learner can recover the right cell and cite the right page is a separate question. Account access and repeated attempts also affect the practical cost.",
      "try_this": "Give Neon two nearly identical manual pages and a small inventory table. Ask for the part compatible with revision B, its evidence location and the rejected alternative. Try a second effort level on the same materials.",
      "sources": [
        "gemini-neon",
        "aa-gemini"
      ]
    },
    {
      "id": "grok47-tools",
      "section": "Models and prompting",
      "provider": "xAI / SpaceXSI",
      "date": "2026-09-21",
      "evidence": "Documented release",
      "availability": "Standard Grok 4.7 is documented for the API. The Fast variant was offered in Cursor and Grok Build.",
      "title": "Grok 4.7 and its Fast option depend on where you run them",
      "change": "The September release and API log document text and image input, text output, several reasoning-effort settings, and a 500,000-token context window for Grok 4.7. The price changes for very long prompts. The Fast variant has its own availability and pricing.",
      "implication": "A speed result reported inside a coding application needs the product and mode recorded beside it. A learner using the public API needs instructions for the available option, especially if the class compares completion time.",
      "try_this": "Use a screenshot of a fictional inventory table. Ask for the missing item and its row number. Compare elapsed time only when the input, requested output and serving mode have been recorded.",
      "sources": [
        "grok-model",
        "grok-log"
      ]
    },
    {
      "id": "chatgpt-live-dictation",
      "section": "Tools and access",
      "provider": "OpenAI",
      "date": "2026-11-18",
      "evidence": "Documented rollout",
      "availability": "Rolling out across plans on iOS, Android, web and desktop.",
      "title": "Dictate a prompt and catch a transcription error before sending",
      "change": "Live Dictation turns speech into visible text while a person is speaking. OpenAI also removed the former ten-minute recording limit. The waveform control reveals or hides the transcript. This is a way to compose a message that a learner can inspect.",
      "implication": "This is an immediate access improvement for a learner who finds typing difficult. In technical work even a fluent transcript can change a unit, a part number or the word “micro.” A short check before submission supports independent participation.",
      "try_this": "Read out “Part M6, two washers, length 15 millimeters.” Inspect the visible text for all three facts, correct it if needed and then send. Offer the same exercise by typing or pasting the sentence.",
      "sources": [
        "chatgpt"
      ]
    },
    {
      "id": "gemini-android-go",
      "section": "Tools and access",
      "provider": "Google",
      "date": "2026-11-18",
      "evidence": "Documented rollout",
      "availability": "Android Go devices with at least 2 GB RAM, an eligible Google account, and supported language and country.",
      "title": "An entry-level Android phone can be a route into the same activity",
      "change": "Google extended Gemini to eligible lower-memory Android Go devices. The help page describes access through the Google app and a supported button gesture. It also lists two absent functions on this edition, the wake phrase and asking about the current screen.",
      "implication": "A phone-based guide can start at an entry point that actually exists. The same written outcome can be achieved with a typed description when the device cannot share its screen. School-account access and family settings can also change the path.",
      "try_this": "Ask the learner to summarize a three-line public notice. Provide the notice as text and let them record which input method and app worked. The score is the accuracy of the summary.",
      "sources": [
        "gemini-mobile",
        "gemini-go"
      ]
    },
    {
      "id": "notebook-source-study",
      "section": "Tools and access",
      "provider": "Google",
      "date": "2026-09-15",
      "evidence": "Documented rollout",
      "availability": "NotebookLM was renamed Gemini Notebook in July. September study features had feature-specific rollouts.",
      "title": "Gemini Notebook can turn a source packet into study practice",
      "change": "Google added interactive learning overviews that can combine summaries, quizzes and flashcards in a notebook. The source collection gives an instructor a practical way to constrain the topic and inspect a generated answer against the reading.",
      "implication": "An AI-generated quiz can help learners rehearse. It can also repeat an error from its own generated key. Keep a human-authored answer key and a fresh independent task for competency verification. An older walkthrough may use the NotebookLM name.",
      "try_this": "Supply a public one-page specification. Have the notebook generate three questions. Ask the learner to underline each answer in the original before trying a fourth question with the notebook closed.",
      "sources": [
        "gemini-notebook",
        "gemini-study"
      ]
    },
    {
      "id": "copilot-mai-code2",
      "section": "Tools and access",
      "provider": "GitHub / Microsoft",
      "date": "2026-11-17",
      "evidence": "Documented release",
      "availability": "MAI Code 2 Flash is listed for Free and Student as well as paid plans. Full Code 2 has a different plan list.",
      "title": "GitHub Copilot offers a coding model that a Free or Student learner can select",
      "change": "GitHub added MAI Code 2 and Code 2 Flash across its supported editors and clients, including GitHub Mobile on iOS and Android. GitHub describes Flash as suited to explanations and focused edits. Organization policy and usage limits still apply.",
      "implication": "For a student working from a phone, explaining an eight-line function and predicting its output is a realistic first activity. It does not require a multi-file agent or an upgrade. Distinguish GitHub Copilot from Microsoft 365 Copilot in navigation instructions.",
      "try_this": "Ask Flash to explain an off-by-one error in a short loop. Before revealing the explanation, have the learner write the expected number of iterations and compare it with the actual result.",
      "sources": [
        "github-mai"
      ]
    },
    {
      "id": "office-copilot-status",
      "section": "Tools and access",
      "provider": "Microsoft",
      "date": "2026-11-17",
      "evidence": "Mixed rollout",
      "availability": "Microsoft 365 work context. Office in Copilot was generally available at announcement. Autopilot began a Frontier rollout.",
      "title": "Office in Copilot and Autopilot have different availability",
      "change": "At Ignite, Microsoft placed Home and Office in Copilot in general availability. Copilot Code had a future general-availability date and Autopilot was starting with Frontier customers. The announcement describes Word, Excel and PowerPoint work within the Copilot experience.",
      "implication": "A work or school lesson can use a small spreadsheet to teach review of AI-generated work. A preview-only control is a poor prerequisite for the same competency. A managed tenant may expose different features from a personal account.",
      "try_this": "Use a fictional supplies sheet. Have the learner check the total and the formula in the resulting spreadsheet. Provide a plain copy of the table for anyone using another spreadsheet app.",
      "sources": [
        "ms-office"
      ]
    },
    {
      "id": "deepseek-hosted-tools",
      "section": "Tools and access",
      "provider": "DeepSeek",
      "date": "2026-11-13",
      "evidence": "Documented release",
      "availability": "DeepSeek V4.1-Flash through its hosted API. Tool charges can add to model-token fees.",
      "title": "DeepSeek brings hosted search and file handling to its API",
      "change": "The November update exposes web search, multi-step research search, document input and URL reading on the hosted service. The provider handles tool calls and prepares the context. A local program needs its own retrieval tools to reproduce that workflow.",
      "implication": "A student comparing a local answer and an online answer may be measuring access to new documents as much as reasoning ability. A fair comparison records which system could retrieve a source and whether web access added a fee.",
      "try_this": "Run a fixed release-note question with the source text supplied to both systems. In a separate trial allow web search. Score whether the answer supports its release date with a matching passage.",
      "sources": [
        "deepseek-api"
      ]
    },
    {
      "id": "grok-citation-audit",
      "section": "Tools and access",
      "provider": "xAI / SpaceXSI",
      "date": null,
      "evidence": "Documented behavior",
      "availability": "Grok developer tools with web or X search. Response shape and citation defaults depend on the API or SDK.",
      "title": "A search-result URL and a citation for a specific claim carry different evidence",
      "change": "The citation docs distinguish URLs encountered by search from inline references attached to a statement. Enabling inline citations still leaves their placement to the model. A long source list can therefore accompany a claim that has little support.",
      "implication": "For fact-checking, ask a learner to match one sentence to one source passage. A reviewer can check the recorded match directly. The same habit transfers to any research assistant.",
      "try_this": "Ask for the release date of a feature and the first platform that supported it. Open the cited link for each claim, inspect the date and label any unsupported claim for follow-up.",
      "sources": [
        "grok-citations"
      ]
    },
    {
      "id": "genie4-worlds",
      "section": "Tools and access",
      "provider": "Google DeepMind",
      "date": "2026-11-18",
      "evidence": "Research preview",
      "availability": "Rollout to eligible Google AI Ultra subscribers. At announcement, Pro members could join a waitlist.",
      "title": "Genie 4 lets a creator change a world while exploring it",
      "change": "Project Genie added mid-session text edits and generated audio that adapts as the world changes. Google lists continuing limitations in physical interactions, audio precision, latency and persistence of a scene.",
      "implication": "This makes a useful optional example of iterative creative direction. A convincing virtual workbench can still change the position of a dial or behave inconsistently between turns. A still storyboard offers a low-cost way to practice the same communication skill.",
      "try_this": "Sketch a room with a clock and two doors. Ask to change only the lighting, then compare the door count and the clock position across frames. Score continuity independently from visual appeal.",
      "sources": [
        "genie"
      ]
    },
    {
      "id": "liquid-small-local",
      "section": "Local and open-weight AI",
      "provider": "Liquid AI",
      "date": "2026-10-27",
      "evidence": "Documented model files",
      "availability": "LFM3-1.7B has downloadable weights. The publisher lists its Q4_K_M GGUF at 1.1 GB and the BF16 file at 3.47 GB.",
      "title": "A 1.1 GB model file makes a modest local experiment concrete",
      "change": "Liquid AI released a small text model alongside other LFM3 variants. Its file list lets a learner inspect the actual download before choosing a build. Quantization stores weights at lower precision. Memory used while running also depends on the runtime and how much conversation it keeps.",
      "implication": "A desktop with 12 GB of graphics memory and 16 GB of system RAM is a reasonable candidate for a bounded test of this small model. Compatibility, speed and answer quality on that machine still need measurement. The text model is only one component of a voice assistant.",
      "try_this": "Try a short, fictional command classification task with a small context. Record exact filename, runtime version, peak memory, first-response delay and correct labels. An entirely paper-based classification is the reference answer.",
      "sources": [
        "liquid-launch",
        "liquid-files"
      ]
    },
    {
      "id": "qwen-thinking-controls",
      "section": "Local and open-weight AI",
      "provider": "Qwen",
      "date": "2026-08-14",
      "evidence": "Documented model files",
      "availability": "Qwen3.8-27B weights and configuration are published. Local inference needs a compatible runtime and adequate memory.",
      "title": "Qwen’s model card exposes settings that can change a comparison",
      "change": "The Qwen team lists an August release for its 27B vision-language model. Its card documents image and video understanding, thinking enabled by default, a per-request switch and reasoning-effort control. These settings make a bare model name an incomplete description of a run.",
      "implication": "A learner comparing a desktop result to a cloud result should log the exact model file, format, context length and thinking setting. If memory is limited, using a smaller member of the family is a different experiment and should get a separate record.",
      "try_this": "Read the card and fill in an experiment sheet before downloading anything. List the model family, parameter count, input types and reasoning setting, then identify which facts still depend on the local runtime.",
      "sources": [
        "qwen-list",
        "qwen-card"
      ]
    },
    {
      "id": "mistral-release-stages",
      "section": "Local and open-weight AI",
      "provider": "Mistral AI",
      "date": "2026-10-29",
      "evidence": "Documented model files",
      "availability": "The API preview was announced October 6. Downloadable weights followed October 29 under the Mistral Ecosystem Open-Weight License.",
      "title": "Mistral Large 4 had a preview date and a later weights release",
      "change": "The later release is a one-trillion-parameter mixture-of-experts model. A mixture-of-experts model uses a subset of its components on a given token. The full set of weight files still matters for storage and deployment planning. The later release also followed further training.",
      "implication": "The archive needs separate entries for an API preview, the actual weights and any measured result. A lab can study the card and compare deployment requirements without buying hardware or presenting a server-scale model as a laptop setup.",
      "try_this": "Make a two-date timeline. Link each date to the artifact a learner could obtain then. Ask which evidence would establish that an October preview benchmark still applies to the later weights.",
      "sources": [
        "mistral-preview",
        "mistral-weights",
        "mistral-hub"
      ]
    },
    {
      "id": "meta-muse-access",
      "section": "Local and open-weight AI",
      "provider": "Meta",
      "date": "2026-10-15",
      "evidence": "Documented model files",
      "availability": "Meta publishes the Muse Spark Open 816B-A42B card with an access form and the Muse Model License.",
      "title": "A public model card can describe a gated download",
      "change": "The card gives the release date, model identity and conditions for access. Someone can inspect those details before requesting the files. The full checkpoint represents far more weights than the subset activated during an individual step.",
      "implication": "“Open weights” by itself tells a student little about account requirements, hardware, the exact license or the feasibility of self-hosting. The useful beginner skill is to find each of those facts and trace it to the publisher.",
      "try_this": "Compare the visible Meta card with Liquid AI’s small GGUF file listing. Record download access, model identity, file information and the license name. Mark missing values explicitly.",
      "sources": [
        "meta-weights",
        "liquid-files"
      ]
    },
    {
      "id": "perplexity-local-mac",
      "section": "Local and open-weight AI",
      "provider": "Perplexity",
      "date": "2026-11-06",
      "evidence": "Documented release",
      "availability": "Launch guidance lists Pro, Max and Enterprise, Mac app 26.42.0 or later, and recommends 36 GB of unified memory.",
      "title": "Perplexity Computer added a fully local mode on supported Macs",
      "change": "The vendor describes an execution mode that can use local files and Mac applications offline when permissions allow. Its product distinguishes local from hybrid execution. Ollama’s own documentation also distinguishes local execution from requests to cloud-hosted models.",
      "implication": "The device showing a chat window and the device doing the inference can be different. For the Academy’s Windows or Linux path, document the chosen runtime separately. The Mac product’s requirements are specific to its implementation.",
      "try_this": "Use three cards labeled interface, model and tools. Draw where a sample note travels for a local-only run and a hosted run. On a device you already own, test the applicable offline path after initial downloads.",
      "sources": [
        "perplex-local",
        "ollama-local"
      ]
    },
    {
      "id": "ssd-expert-streaming",
      "section": "Local and open-weight AI",
      "provider": "Hugging Face",
      "date": "2026-11-18",
      "evidence": "Author experiment",
      "availability": "Engineering demonstration with custom components on an Apple M5 Mac. Academy hardware reproduction is open.",
      "title": "An SSD-streaming demo worked on a 32 GB Mac under specific conditions",
      "change": "The author ran a roughly 180B mixture-of-experts model by keeping some weights on disk and fetching selected blocks as needed. One reported configuration peaked at 25.4 GiB with a 6,000-token input. The post describes quantization, cache settings and the hardware used.",
      "implication": "This is evidence that the technique worked in that setup. A Windows PC with separate GPU and system memory has a different memory path. Adding the two headline memory figures does not predict equivalent performance. No purchase is needed to learn from the experiment.",
      "try_this": "Keep this on the watchlist. Evidence for a Windows lab would include a supported build, exact hardware, identical prompts, measured memory and usable response times on repeated runs.",
      "sources": [
        "stream-demo"
      ]
    },
    {
      "id": "aa-gpt62-split",
      "section": "Performance evidence",
      "provider": "Artificial Analysis",
      "date": "2026-11-17",
      "evidence": "Independent evaluation",
      "availability": "Independent launch evaluation. The coding comparison used Codex at maximum reasoning effort.",
      "title": "GPT-6.2’s broad gains did not carry over to every coding test",
      "change": "Artificial Analysis measured a five-point general Intelligence Index gain for Astra over GPT-6 Astra. On its separate Coding Agent Index, Astra was level with its predecessor and GPT-6.2 Sol scored three points below GPT-6.1 Sol. The evaluator traced part of Sol’s drop to omitted facts on repository questions.",
      "implication": "A single “best model for coding” line would hide the distinction between solving a well-specified task and answering questions about an existing project. This is one evaluator’s result under a named setup. It provides a reason to test the Academy’s own assignments.",
      "try_this": "Test explanation and repair separately on a small program. Count correct facts in the explanation, then run an independent check on the repair. Preserve both results even when they point to different choices.",
      "sources": [
        "aa-openai"
      ]
    },
    {
      "id": "aa-neon-carbon",
      "section": "Performance evidence",
      "provider": "Artificial Analysis",
      "date": "2026-11-18",
      "evidence": "Independent evaluation",
      "availability": "Neon at xhigh effort versus Carbon at high. Values describe this evaluator’s task mix.",
      "title": "Neon’s cost comparison changes with the price used for Carbon",
      "change": "Artificial Analysis reported index scores of 61 for Neon and 60 for Carbon. Its mean task cost was $6.44 for Neon and $7.76 for Carbon using Carbon’s standard rates. Carbon also had a temporary price discount, so the 17% saving refers to the stated standard-price comparison.",
      "implication": "Two screenshots of a leaderboard can both be accurate while using different cost assumptions. For a learner’s budget, the relevant denominator is a successfully completed task, including retries. A one-point score difference on the index is insufficient evidence for replacing a guide.",
      "try_this": "Put price date, effort level, success definition and retry count next to every quoted cost. Recalculate a small comparison with both standard and discounted prices to see whether the ordering changes.",
      "sources": [
        "aa-gemini",
        "gemini-neon"
      ]
    },
    {
      "id": "ms-mai-cost-claim",
      "section": "Performance evidence",
      "provider": "Microsoft AI",
      "date": "2026-11-17",
      "evidence": "Provider evaluation",
      "availability": "Publisher comparison of MAI Code 2 with GPT-6.1 Sol, using Microsoft’s selected benchmark mix.",
      "title": "Microsoft’s 98% figure is a relative score on four coding evaluations",
      "change": "Microsoft reports that MAI Code 2 averaged 98% of GPT-6.1 Sol’s performance at 40% of the cost across four engineering benchmarks. These covered coding tasks and included named versions of DeepSWE, Terminal-Bench, SWE-Atlas and FrontierCode. The 98% is relative performance against a comparator.",
      "implication": "Reading that figure as a 98% success rate on student questions changes its meaning. It is reasonable evidence for testing a cheaper coding model. The claimed percentages should keep their attribution until independently reproduced on a comparable workload.",
      "try_this": "Show a hypothetical reference score of 50 and ask what 98% of it would be. The answer is 49. Next, ask what additional data would establish how often the model correctly explains a beginner’s loop.",
      "sources": [
        "ms-mai",
        "github-mai"
      ]
    },
    {
      "id": "metr-time-horizon",
      "section": "Performance evidence",
      "provider": "METR",
      "date": null,
      "evidence": "Evaluation method",
      "availability": "Historical independent methodology. METR labels this chart as no longer actively updated.",
      "title": "An agent’s time horizon measures the human length of a task",
      "change": "METR fitted success probabilities to tasks timed by human experts. A 50% horizon is the human task duration at which the agent is predicted to succeed half the time on that distribution. The tasks were largely well-specified technical work. METR flags estimates above 16 hours as unreliable in that suite.",
      "implication": "A two-hour horizon is an unreliable basis for an expectation of two uninterrupted hours of autonomous classroom work. A more meaningful beginner exercise checks whether the model completed a bounded job and whether the learner can independently verify the outcome.",
      "try_this": "Give a software assistant a clearly scoped task with three acceptance conditions, a deadline and a fixture. Record pass or fail on each condition and elapsed machine time as a separate measurement.",
      "sources": [
        "metr-method"
      ]
    }
  ],
  "exercises": [
    {
      "title": "Find the current revision",
      "material": "Two fictional records. Bulletin A, July 8: “Logger P records every 30 seconds.” Bulletin B, August 3: “Revision B of Logger P records every 20 seconds. Revision A stays at 30 seconds.”",
      "prompt": "For revision B, give the interval, identify the supporting bulletin, and say whether the records specify battery life. Use only these records.",
      "answer": "20 seconds for revision B, supported by Bulletin B dated August 3. Battery life is unspecified.",
      "score": "Award one point each for the interval, the correct bulletin and the missing-information judgment. A fabricated battery-life figure loses the third point. Accessible path: read, type or dictate the same two records."
    },
    {
      "title": "Audit a parts total",
      "material": "A fictional parts slip lists two spacers at $4.50 each, three washers at $1.20 each, and one label roll at $6.25. The printed subtotal is $19.85. There is no tax or shipping.",
      "prompt": "Check the three line totals and the subtotal. If you find a discrepancy, give its size and show the arithmetic.",
      "answer": "$9.00 + $3.60 + $6.25 = $18.85. The printed subtotal is $1.00 too high.",
      "score": "Award one point for all line totals, one for the corrected subtotal and one for the direction and size of the discrepancy. Try a plain prompt first, then the same task with a spreadsheet tool. Label the tool-enabled run separately."
    },
    {
      "title": "Explain and repair a boundary error",
      "material": "A small Python function is intended to return readings strictly above a limit. Its body is “return [x for x in values if x >= limit]”. The sample readings are [20, 22, 31, 19] and the limit is 22.",
      "prompt": "Predict the current output, explain the mismatch in one sentence, and suggest the smallest fix. Give one test that would catch the original error.",
      "answer": "The current output is [22, 31]. The intended output is [31]. The comparison should use >. With readings [22] and limit 22, the expected result is an empty list.",
      "score": "Award one point for the actual output, one for the correct fix and one for a test that fails on the original code. Have the learner explain the boundary in their own words before considering the skill independently demonstrated."
    },
    {
      "title": "Route desk-robot requests without hardware",
      "material": "Use these labels: TIMER, NOTE, CLARIFY. The printed cards say “Set a timer for five minutes”, “Remember that the blue lead is in drawer B”, and “Do the usual thing”. Treat the third message as missing the intended action.",
      "prompt": "Choose one allowed label for each card. Give one short question for any card labeled CLARIFY. Return labels in the same order.",
      "answer": "TIMER, NOTE, CLARIFY. A suitable question asks which action the speaker wants. No device action is needed.",
      "score": "Record correct labels out of three, whether the assistant invented an action for the ambiguous card, and response time if measured. Use the same cards for a local model and an already available hosted model. A paper sort supplies the control."
    }
  ],
  "sources": {
    "openai-release": {
      "publisher": "OpenAI",
      "title": "Introducing GPT-6.2 Astra and Sol",
      "url": "https://openai.com/index/introducing-gpt-6-2-astra-sol/",
      "published": "2026-11-17",
      "kind": "Release"
    },
    "openai-api": {
      "publisher": "OpenAI",
      "title": "Model release notes • GPT-6.2 API entry",
      "url": "https://help.openai.com/en/articles/9624314-model-release-notes",
      "published": "2026-11-17",
      "kind": "Release log"
    },
    "chatgpt": {
      "publisher": "OpenAI",
      "title": "ChatGPT release notes • November 17–18 entries",
      "url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes",
      "published": null,
      "kind": "Product log"
    },
    "openai-migration": {
      "publisher": "OpenAI",
      "title": "Using GPT-6.2 • Responses API and prompting",
      "url": "https://developers.openai.com/api/docs/guides/latest-model",
      "published": null,
      "kind": "Technical guide"
    },
    "claude-release": {
      "publisher": "Anthropic",
      "title": "Introducing Claude Fable 5.2 and Claude Mythos 5.2",
      "url": "https://www.anthropic.com/claude-fable-and-mythos-5-2",
      "published": "2026-11-09",
      "kind": "Release"
    },
    "claude-migration": {
      "publisher": "Anthropic",
      "title": "Fable 5.2 and Mythos 5.2 migration guide",
      "url": "https://platform.claude.com/docs/en/models/fable-5-2/migration-guide",
      "published": null,
      "kind": "Technical guide"
    },
    "claude-small": {
      "publisher": "Anthropic",
      "title": "Introducing Claude Haiku 5.5",
      "url": "https://www.anthropic.com/claude-haiku-5-5",
      "published": "2026-10-07",
      "kind": "Release"
    },
    "gemini-neon": {
      "publisher": "Google",
      "title": "Introducing Gemini 4 Neon",
      "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/",
      "published": "2026-11-18",
      "kind": "Release"
    },
    "gemini-mobile": {
      "publisher": "Google",
      "title": "Gemini on lower-memory Android devices",
      "url": "https://blog.google/innovation-and-ai/products/gemini-app/gemini-lower-memory-devices/",
      "published": "2026-11-18",
      "kind": "Release"
    },
    "gemini-go": {
      "publisher": "Google",
      "title": "Gemini on Android Go • eligibility and limitations",
      "url": "https://support.google.com/gemini/answer/17079170",
      "published": null,
      "kind": "Product help"
    },
    "gemini-notebook": {
      "publisher": "Google",
      "title": "NotebookLM is now Gemini Notebook",
      "url": "https://blog.google/innovation-and-ai/products/gemini-notebook/notebooklm-gemini-notebook/",
      "published": "2026-07-16",
      "kind": "Release"
    },
    "gemini-study": {
      "publisher": "Google",
      "title": "New Gemini Notebook study tools",
      "url": "https://blog.google/innovation-and-ai/products/gemini-notebook/new-study-tools-september-2026/",
      "published": "2026-09-15",
      "kind": "Release"
    },
    "genie": {
      "publisher": "Google DeepMind",
      "title": "Project Genie update featuring Genie 4",
      "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/genie-4-now-available-project-genie/",
      "published": "2026-11-18",
      "kind": "Release"
    },
    "grok-model": {
      "publisher": "xAI / SpaceXSI",
      "title": "Introducing Grok 4.7",
      "url": "https://x.ai/news/grok-4-7",
      "published": "2026-09-21",
      "kind": "Release"
    },
    "grok-log": {
      "publisher": "xAI / SpaceXSI",
      "title": "API release notes • September through November entries",
      "url": "https://docs.x.ai/developers/release-notes",
      "published": null,
      "kind": "Release log"
    },
    "grok-citations": {
      "publisher": "xAI / SpaceXSI",
      "title": "Citations in tool-assisted responses",
      "url": "https://docs.x.ai/developers/tools/citations",
      "published": null,
      "kind": "Technical guide"
    },
    "github-mai": {
      "publisher": "GitHub",
      "title": "MAI Code 2 and MAI Code 2 Flash in GitHub Copilot",
      "url": "https://github.blog/changelog/2026-11-17-mai-code-2-and-mai-code-2-flash-in-github-copilot/",
      "published": "2026-11-17",
      "kind": "Product log"
    },
    "ms-mai": {
      "publisher": "Microsoft AI",
      "title": "Introducing MAI Code 2 and Code 2 Flash",
      "url": "https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/",
      "published": "2026-11-17",
      "kind": "Release and provider evaluation"
    },
    "ms-office": {
      "publisher": "Microsoft",
      "title": "Copilot at Microsoft Ignite • Office, Code and Autopilot",
      "url": "https://www.microsoft.com/en-us/copilot/blog/2026/11/17/copilot-at-microsoft-ignite-expanding-whats-possible-with-intelligence-built-for-work/",
      "published": "2026-11-17",
      "kind": "Release"
    },
    "deepseek-api": {
      "publisher": "DeepSeek",
      "title": "The full DeepSeek App experience, now via API",
      "url": "https://api-docs.deepseek.com/news/news261113/",
      "published": "2026-11-13",
      "kind": "Release"
    },
    "perplex-local": {
      "publisher": "Perplexity",
      "title": "Fully local Computer on Mac and Voice mode",
      "url": "https://www.perplexity.ai/changelog/teammates-fully-local-computer-on-mac-and-voice-mode",
      "published": "2026-11-06",
      "kind": "Product log"
    },
    "liquid-launch": {
      "publisher": "Liquid AI",
      "title": "Introducing LFM3",
      "url": "https://www.liquid.ai/blog/lfm3",
      "published": "2026-10-27",
      "kind": "Release"
    },
    "liquid-files": {
      "publisher": "Liquid AI",
      "title": "LFM3-1.7B GGUF files and download sizes",
      "url": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF/tree/main",
      "published": null,
      "kind": "Model files"
    },
    "qwen-list": {
      "publisher": "Qwen team",
      "title": "Qwen3.8 repository • release history",
      "url": "https://github.com/QwenLM/Qwen3.8",
      "published": null,
      "kind": "Release log"
    },
    "qwen-card": {
      "publisher": "Qwen team",
      "title": "Qwen3.8-27B model card",
      "url": "https://huggingface.co/Qwen/Qwen3.8-27B",
      "published": null,
      "kind": "Model card"
    },
    "mistral-preview": {
      "publisher": "Mistral AI",
      "title": "Introducing Mistral Large 4 • public preview",
      "url": "https://mistral.ai/news/mistral-large-4/",
      "published": "2026-10-06",
      "kind": "Release"
    },
    "mistral-weights": {
      "publisher": "Mistral AI",
      "title": "Mistral Large 4 weights release",
      "url": "https://mistral.ai/news/mistral-large-4-oss/",
      "published": "2026-10-29",
      "kind": "Release"
    },
    "mistral-hub": {
      "publisher": "Hugging Face",
      "title": "Welcome Le Chonk • files, license and serving",
      "url": "https://huggingface.co/blog/mistral-large4",
      "published": "2026-10-29",
      "kind": "Technical overview"
    },
    "meta-weights": {
      "publisher": "Meta",
      "title": "Muse Spark Open 816B-A42B • model card and license",
      "url": "https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B",
      "published": "2026-10-15",
      "kind": "Model card"
    },
    "stream-demo": {
      "publisher": "Hugging Face engineering blog",
      "title": "Streaming expert weights on a 32 GB Mac",
      "url": "https://huggingface.co/blog/moe-expert-streaming",
      "published": "2026-11-18",
      "kind": "Author experiment"
    },
    "ollama-local": {
      "publisher": "Ollama",
      "title": "Local FAQ • local and cloud execution",
      "url": "https://docs.ollama.com/faq",
      "published": null,
      "kind": "Technical guide"
    },
    "aa-openai": {
      "publisher": "Artificial Analysis",
      "title": "GPT-6.2 release evaluation • general and coding indices",
      "url": "https://artificialanalysis.ai/articles/openai-has-the-2-model-in-the-intelligence-index-with-the-release-of-gpt-6-2-astra-and-gpt-6-2-sol",
      "published": "2026-11-17",
      "kind": "Independent evaluation"
    },
    "aa-gemini": {
      "publisher": "Artificial Analysis",
      "title": "Gemini 4 Neon • quality, cost and time per task",
      "url": "https://artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain",
      "published": "2026-11-18",
      "kind": "Independent evaluation"
    },
    "metr-method": {
      "publisher": "METR",
      "title": "Task-completion time horizons • archived methodology",
      "url": "https://metr.org/time-horizons/",
      "published": null,
      "kind": "Independent methodology"
    }
  }
}
