Tech Adventure Academy

Academy AI Review / Kickoff dossier / Nov 19, 2026

Which AI changes should change a lesson?

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.

What matters for learners

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.

[3] OpenAI [10] Google [9] Google

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.

[32] Artificial Analysis [33] Artificial Analysis

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.

[4] OpenAI [6] Anthropic

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.

[23] Liquid AI [30] Hugging Face engineering blog

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.

Research findings (25)

Models and prompting

OpenAIDocumented release

Where learners can actually select GPT-6.2

In reporting window · Nov 17, 2026

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.

Access and scope. API release and desktop preview in Work and Codex on eligible plans. Managed seats depend on workspace settings.

Why it matters. 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 it. 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.

[1] OpenAI [2] OpenAI [3] OpenAI

OpenAIDocumented behavior

Reusable instructions have a new home in the Responses API

In reporting window · Nov 17, 2026

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.

Access and scope. Advanced track for API users. The behavior concerns the request sent by an application.

Why it matters. 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 it. 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.

[4] OpenAI [1] OpenAI

AnthropicDocumented release

One underlying Claude model has two access paths

Earlier baseline · Nov 9, 2026

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.

Access and scope. Fable is offered on the Claude Platform and supported cloud platforms. Mythos requires organizational verification.

Why it matters. 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 it. 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.

[5] Anthropic

AnthropicDocumented behavior

A shorter answer and a smaller reasoning budget are different requests

Earlier baseline · Nov 9, 2026

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.

Access and scope. Applies to Fable 5.2 API migration. Effort controls exposed by other products can differ.

Why it matters. 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 it. 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.

[6] Anthropic [5] Anthropic

AnthropicProvider positioning

A smaller Claude model gives a routing experiment a starting point

Earlier baseline · Oct 7, 2026

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.

Access and scope. Available as claude-haiku-5-5 on the Claude Platform and through supported cloud providers.

Why it matters. 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 it. 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.

[7] Anthropic

GoogleDocumented release

Gemini 4 Neon adds a middle choice for complex work

In reporting window · Nov 18, 2026

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.

Access and scope. Paid Gemini API tier. The Gemini app and Antigravity rollout includes eligible Google AI Pro and Ultra subscribers.

Why it matters. 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 it. 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.

[8] Google [33] Artificial Analysis

xAI / SpaceXSIDocumented release

Grok 4.7 and its Fast option depend on where you run them

Earlier baseline · Sep 21, 2026

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.

Access and scope. Standard Grok 4.7 is documented for the API. The Fast variant was offered in Cursor and Grok Build.

Why it matters. 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 it. 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.

[14] xAI / SpaceXSI [15] xAI / SpaceXSI

Tools and access

OpenAIDocumented rollout

Dictate a prompt and catch a transcription error before sending

In reporting window · Nov 18, 2026

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.

Access and scope. Rolling out across plans on iOS, Android, web and desktop.

Why it matters. 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 it. 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.

[3] OpenAI

GoogleDocumented rollout

An entry-level Android phone can be a route into the same activity

In reporting window · Nov 18, 2026

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.

Access and scope. Android Go devices with at least 2 GB RAM, an eligible Google account, and supported language and country.

Why it matters. 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 it. 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.

[9] Google [10] Google

GoogleDocumented rollout

Gemini Notebook can turn a source packet into study practice

Earlier baseline · Sep 15, 2026

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.

Access and scope. NotebookLM was renamed Gemini Notebook in July. September study features had feature-specific rollouts.

Why it matters. 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 it. 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.

[11] Google [12] Google

GitHub / MicrosoftDocumented release

GitHub Copilot offers a coding model that a Free or Student learner can select

In reporting window · Nov 17, 2026

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.

Access and scope. MAI Code 2 Flash is listed for Free and Student as well as paid plans. Full Code 2 has a different plan list.

Why it matters. 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 it. 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.

[17] GitHub

MicrosoftMixed rollout

Office in Copilot and Autopilot have different availability

In reporting window · Nov 17, 2026

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.

Access and scope. Microsoft 365 work context. Office in Copilot was generally available at announcement. Autopilot began a Frontier rollout.

Why it matters. 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 it. 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.

[19] Microsoft

DeepSeekDocumented release

DeepSeek brings hosted search and file handling to its API

In reporting window · Nov 13, 2026

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.

Access and scope. DeepSeek V4.1-Flash through its hosted API. Tool charges can add to model-token fees.

Why it matters. 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 it. 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.

[20] DeepSeek

xAI / SpaceXSIDocumented behavior

A search-result URL and a citation for a specific claim carry different evidence

Reference checked Nov 19, 2026

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.

Access and scope. Grok developer tools with web or X search. Response shape and citation defaults depend on the API or SDK.

Why it matters. 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 it. 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.

[16] xAI / SpaceXSI

Google DeepMindResearch preview

Genie 4 lets a creator change a world while exploring it

In reporting window · Nov 18, 2026

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.

Access and scope. Rollout to eligible Google AI Ultra subscribers. At announcement, Pro members could join a waitlist.

Why it matters. 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 it. 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.

[13] Google DeepMind

Local and open-weight AI

Liquid AIDocumented model files

A 1.1 GB model file makes a modest local experiment concrete

Earlier baseline · Oct 27, 2026

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.

Access and scope. 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.

Why it matters. 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 it. 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.

[22] Liquid AI [23] Liquid AI

QwenDocumented model files

Qwen’s model card exposes settings that can change a comparison

Earlier baseline · Aug 14, 2026

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.

Access and scope. Qwen3.8-27B weights and configuration are published. Local inference needs a compatible runtime and adequate memory.

Why it matters. 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 it. 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.

[24] Qwen team [25] Qwen team

Mistral AIDocumented model files

Mistral Large 4 had a preview date and a later weights release

Earlier baseline · Oct 29, 2026

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.

Access and scope. The API preview was announced October 6. Downloadable weights followed October 29 under the Mistral Ecosystem Open-Weight License.

Why it matters. 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 it. 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.

[26] Mistral AI [27] Mistral AI [28] Hugging Face

MetaDocumented model files

A public model card can describe a gated download

Earlier baseline · Oct 15, 2026

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.

Access and scope. Meta publishes the Muse Spark Open 816B-A42B card with an access form and the Muse Model License.

Why it matters. “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 it. 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.

[29] Meta [23] Liquid AI

PerplexityDocumented release

Perplexity Computer added a fully local mode on supported Macs

Earlier baseline · Nov 6, 2026

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.

Access and scope. Launch guidance lists Pro, Max and Enterprise, Mac app 26.42.0 or later, and recommends 36 GB of unified memory.

Why it matters. 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 it. 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.

[21] Perplexity [31] Ollama

Hugging FaceAuthor experiment

An SSD-streaming demo worked on a 32 GB Mac under specific conditions

In reporting window · Nov 18, 2026

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.

Access and scope. Engineering demonstration with custom components on an Apple M5 Mac. Academy hardware reproduction is open.

Why it matters. 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 it. 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.

[30] Hugging Face engineering blog

Performance evidence

Artificial AnalysisIndependent evaluation

GPT-6.2’s broad gains did not carry over to every coding test

In reporting window · Nov 17, 2026

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.

Access and scope. Independent launch evaluation. The coding comparison used Codex at maximum reasoning effort.

Why it matters. 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 it. 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.

[32] Artificial Analysis

Artificial AnalysisIndependent evaluation

Neon’s cost comparison changes with the price used for Carbon

In reporting window · Nov 18, 2026

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.

Access and scope. Neon at xhigh effort versus Carbon at high. Values describe this evaluator’s task mix.

Why it matters. 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 it. 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.

[33] Artificial Analysis [8] Google

Microsoft AIProvider evaluation

Microsoft’s 98% figure is a relative score on four coding evaluations

In reporting window · Nov 17, 2026

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.

Access and scope. Publisher comparison of MAI Code 2 with GPT-6.1 Sol, using Microsoft’s selected benchmark mix.

Why it matters. 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 it. 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.

[18] Microsoft AI [17] GitHub

METREvaluation method

An agent’s time horizon measures the human length of a task

Reference checked Nov 19, 2026

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.

Access and scope. Historical independent methodology. METR labels this chart as no longer actively updated.

Why it matters. 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 it. 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.

[34] METR

Practice tasks

Find the current revision

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 key and scoring

20 seconds for revision B, supported by Bulletin B dated August 3. Battery life is unspecified.

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.

Audit a parts total

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 key and scoring

$9.00 + $3.60 + $6.25 = $18.85. The printed subtotal is $1.00 too high.

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.

Explain and repair a boundary error

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 key and scoring

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.

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.

Route desk-robot requests without hardware

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 key and scoring

TIMER, NOTE, CLARIFY. A suitable question asks which action the speaker wants. No device action is needed.

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.

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.

Source index (34)

Sources checked Nov 19, 2026. A living source may change after this edition.

  1. Introducing GPT-6.2 Astra and SolOpenAI · Release · Nov 17, 2026
  2. Model release notes • GPT-6.2 API entryOpenAI · Release log · Nov 17, 2026
  3. ChatGPT release notes • November 17–18 entriesOpenAI · Product log · Living source
  4. Using GPT-6.2 • Responses API and promptingOpenAI · Technical guide · Living source
  5. Introducing Claude Fable 5.2 and Claude Mythos 5.2Anthropic · Release · Nov 9, 2026
  6. Fable 5.2 and Mythos 5.2 migration guideAnthropic · Technical guide · Living source
  7. Introducing Claude Haiku 5.5Anthropic · Release · Oct 7, 2026
  8. Introducing Gemini 4 NeonGoogle · Release · Nov 18, 2026
  9. Gemini on lower-memory Android devicesGoogle · Release · Nov 18, 2026
  10. Gemini on Android Go • eligibility and limitationsGoogle · Product help · Living source
  11. NotebookLM is now Gemini NotebookGoogle · Release · Jul 16, 2026
  12. New Gemini Notebook study toolsGoogle · Release · Sep 15, 2026
  13. Project Genie update featuring Genie 4Google DeepMind · Release · Nov 18, 2026
  14. Introducing Grok 4.7xAI / SpaceXSI · Release · Sep 21, 2026
  15. API release notes • September through November entriesxAI / SpaceXSI · Release log · Living source
  16. Citations in tool-assisted responsesxAI / SpaceXSI · Technical guide · Living source
  17. MAI Code 2 and MAI Code 2 Flash in GitHub CopilotGitHub · Product log · Nov 17, 2026
  18. Introducing MAI Code 2 and Code 2 FlashMicrosoft AI · Release and provider evaluation · Nov 17, 2026
  19. Copilot at Microsoft Ignite • Office, Code and AutopilotMicrosoft · Release · Nov 17, 2026
  20. The full DeepSeek App experience, now via APIDeepSeek · Release · Nov 13, 2026
  21. Fully local Computer on Mac and Voice modePerplexity · Product log · Nov 6, 2026
  22. Introducing LFM3Liquid AI · Release · Oct 27, 2026
  23. LFM3-1.7B GGUF files and download sizesLiquid AI · Model files · Living source
  24. Qwen3.8 repository • release historyQwen team · Release log · Living source
  25. Qwen3.8-27B model cardQwen team · Model card · Living source
  26. Introducing Mistral Large 4 • public previewMistral AI · Release · Oct 6, 2026
  27. Mistral Large 4 weights releaseMistral AI · Release · Oct 29, 2026
  28. Welcome Le Chonk • files, license and servingHugging Face · Technical overview · Oct 29, 2026
  29. Muse Spark Open 816B-A42B • model card and licenseMeta · Model card · Oct 15, 2026
  30. Streaming expert weights on a 32 GB MacHugging Face engineering blog · Author experiment · Nov 18, 2026
  31. Local FAQ • local and cloud executionOllama · Technical guide · Living source
  32. GPT-6.2 release evaluation • general and coding indicesArtificial Analysis · Independent evaluation · Nov 17, 2026
  33. Gemini 4 Neon • quality, cost and time per taskArtificial Analysis · Independent evaluation · Nov 18, 2026
  34. Task-completion time horizons • archived methodologyMETR · Independent methodology · Living source