Source: six YouTube auto-caption transcripts fetched 2026-11-01: raw/New_Leading_AI_Models_as_Benchmarks_Shift.md (The AI Daily Brief, Nathaniel Whittemore “NLW”; sponsors KPMG, CData, Granola, Blitzy), raw/Google_Carbon_just_won_everything....md (Matthew Berman; sponsored by Factory), raw/AI_Is_Moving_Too_Fast._Two_New_Models_In_Two_Weeks.md (AI For Humans, Gavin Purcell and Kevin Pereira), raw/GPT-6.1_Terra_vs_Claude_5.5_-_Same_Prompt_Not_Even_Close.md (MattVidPro), raw/They_actually_did_it..._Kimi_K3.1.md (Matthew Berman; sponsored by CodeRabbit) and raw/Deepseek_crushed_everyone.md (Matthew Berman; sponsored by CoreWeave).
2026-11-09 addition (section “First week of November”): Matthew Berman’s AI news roundup (raw/Kimi_K3.1_weights_OUT_open_source_computer_use_SOTA_image_model_science_model_Pika_H3_-_AI_NEWS.md, https://www.youtube.com/watch?v=-M62vWrZh54), Matt Wolfe (raw/AI_News_-_New_ChatGPT_Google_Models_14_Updates_Worth_Knowing.md, https://www.youtube.com/watch?v=IrAHINf0M6E; sponsored by Optimizely), AI Explained (raw/We_may_release_models_we_can_t_control_-_Amodei_Altman_..._and_Gemini_Carbon.md), The AI Daily Brief (raw/AI_Lurches_Towards_RSI.md) and The Neuron Daily for November 3, 5, 6 and 9 (raw/newsletter-theneurondaily-com-{ee205cba23,ef2052c870,2cee76fc49,27041e78ca}.md), plus GDPX and LifeEval detail from Moonshots EP308 (raw/Big_Tech_s_3T_AI_Bet_Claude_Writes_90%_of_Anthropic_s_Code_AIs_Editing_AI_Brains_EP_308.md) and model rumours from AI For Humans (raw/AI_Just_Lost_The_Election._Bigger_Models_Are_Coming_Anyway..md), applied from other writers’ proposals.
2026-11-11 addition (section “Second week of November”): Matthew Berman’s news roundup (raw/Gemini_Carbon_NVIDIA_CEO_AFFAIR_Fable_5.2_Meta_Muse_Ember_GPT_Image_2.6_Minimax_H3_-_AI_NEWS.md, https://www.youtube.com/watch?v=FCmLuTENOp0; sponsored by Victor) and The Neuron Daily for November 10 and 11 (raw/newsletter-theneurondaily-com-{47e5bd68d8,14c86ec601}.md), checked against first-party pages fetched 2026-11-11: Google’s Carbon post and developer guide, Meta’s Muse Ember post, Thinking Machines’ Inkling 2 Small post and MiniMax’s H3-Image model card (ai-research/ files in frontmatter).
Creator: multiple (see above) · URL: https://www.youtube.com/watch?v=Dz4_db_rRDE (NLW), https://www.youtube.com/watch?v=zhnJo7s91ls (Berman, Carbon), https://www.youtube.com/watch?v=6101Q0dQQIk (AI For Humans), https://www.youtube.com/watch?v=6Il7tudXo1E (MattVidPro), https://www.youtube.com/watch?v=5sEVvu4mQ8g (Berman, Kimi) · Duration: not recorded · Platform: YouTube
This picks up from the October landscape and covers the last week of October. Three things changed the picture: Google announced Gemini 4 Carbon two weeks after Argon and it took first place on the index; Artificial Analysis replaced its Intelligence Index with v5, which reshuffled every ranking and widened the gap between closed and open models; and OpenAI and Moonshot filled in the middle of the market with GPT-6.1 Terra and Kimi K3.1 Pro. Every benchmark below is vendor-reported or an Artificial Analysis (AA) chart read aloud by a creator; nobody fetched AA’s page for this article.
Key Takeaways
- AA Intelligence Index v5 is a new scale. Do not compare its numbers with any earlier AA score. Opus 5.5 is 58 on both versions, but Argon went from 53 to 50, Kimi K3.1 Pro from 55 to 46, and the cost-per-task figures roughly doubled for some models (Fable 5.1 from 13.04). ^[inferred: pre-v5 figures from Berman and the October article, v5 figures from NLW]
- Gemini 4 Carbon leads v5 at 59, one point above Opus 5.5, but you cannot use it yet. Google announced it with benchmarks and a price (10 per million tokens introductory, then 20) and is giving it first to cyber defenders. No date for general access.
- v5 rewards agentic and science work. In NLW’s placings (counted before Carbon was scored), GPT-6 Astra jumped from 6th to 2nd (57) and GPT-6.1 Sol from 9th to 3rd. Sonnet 5.5 slipped from 2nd to 4th and Fable 5.1 from 5th to 8th.
- The open-weight gap widened from 3 points to 12 (AA, via NLW). Kimi K3.1 Pro is the top open model at 46; DeepSeek V4.1 Flash and GLM 5.3 are at 32, below Claude Haiku 5.5 at 36.
- Cost per task now spans 14x inside the top 10: GPT-6.1 Terra 1.27, Astra 7.76, Opus 5.5 and Sonnet 5.5 about 13.04 (AA v5, via NLW).
- GPT-6.1 Terra is a computer-use workhorse, not a creative model. Fifth on v5 before Carbon was added, and the cheapest frontier model per task, but two hands-on testers found it weak at following creative direction and at 3D.
- Kimi K3.1 Pro is the best open model on paper and fabricates when unsupervised. When asked for a benchmark chart it invented the numbers and called them “invented placeholders” only when Berman asked.
- Gemini 4 Hydrogen (added 2026-11-13): Google’s new Flash-tier model replaced Gemini 3.8 Flash across its products on November 12 at the same API price, 7.50 per million tokens. It has four thinking levels (low, medium, high, xhigh), scores 54 on the Artificial Analysis index at xhigh against Argon’s 50, and now powers AI Overviews and AI Mode. Google positions it for simpler code generation and test runs; Argon and Carbon stay the long-horizon models. Ars says free users stay on 3.5 Flash-Lite; Google’s own post says free users get Hydrogen through automatic model selection. ^[ambiguous] (
ai-research/arstechnica-gemini-4-hydrogen-replaces-3-8-flash-2026-11-13.md) - Second week of November (added 2026-11-11): Carbon opened to paying developers on November 10 with Antigravity’s
/swarm; Claude Fable 5.2 took the top of AA’s index; Meta’s Muse Ember priced like Haiku 5.5 at about a third of the cost per task; MiniMax released H3-Image weights. See “Second week of November” below. - First week of November (added 2026-11-09): Kimi K3.1 Pro’s weights shipped; OpenAI released a restricted GPT-6 Cyber; Carbon placed second to Opus 5.5 on SemiAnalysis’s new GDPX benchmark at about a third of the cost per task; a rumoured GPT-6.2 Sol would tie Carbon at the top of AA v5; and a dozen smaller open-weight models landed. See the section below.
At a Glance
| Model (lab) | Status | List price, $/M in / out | AA v5 | Cost per task (AA v5) | Consider it when |
|---|---|---|---|---|---|
| Gemini 4 Carbon (Google) | Announced ~Oct 28 to cyber defenders; open to paying API developers and Ultra subscribers from November 10 (updated 2026-11-11) | 10 through November 30, then 20 from December 1; cached input 95% off | 59 (#1 before Fable 5.2) | $7.76 | Long-horizon agent, research and multi-agent /swarm jobs; Argon for everyday work (Google’s own guidance) |
| GPT-6.1 Terra (OpenAI) | Released late October | About half GPT-6.1 Sol’s price (MattVidPro); exact list price not in these sources | 6th with Carbon, 5th before (score not given) | $0.96 | Long, cheap computer-use loops ^[inferred] |
| Kimi K3.1 Pro (Moonshot) | Kimi app, desktop app and API; open weights on Hugging Face from the first week of November (custom licence; updated 2026-11-09) | 10, 4 / $20 | 46 (#10 with Carbon; top open model) | $2.11 (pre-v5, Berman) | Open-weight coding and terminal work where you can check the output ^[inferred] |
| DeepSeek V4.1 Flash 1017 | Open weights (MIT) | 0.60 off-peak | 32 | $0.26 (pre-v5, Berman) | Cheap bulk work; see DeepSeek V4.1 Flash |
Artificial Analysis Intelligence Index v5
What changed (NLW, reading AA’s announcement):
- Most benchmarks are now private. AA says this is the first version where the majority run on private data sets, to make “benchmark maxing” harder. Kyle Chan (Brookings) puts the private share at 60%, up from 45%.
- Added: Frontier Code 2.0 (Cognition’s private set of real coding tasks, scored on whether a repository maintainer would merge the code) and Terminal-Bench Science (agentic research).
- Made harder: Humanity’s Last Exam becomes HLE Diamond; AA Omniscience (knowledge and hallucination) and a physics test the captions call “Crux PT” get harder versions.
- Replaced: GDPval-AA is out; white-collar work is now measured by AA Briefcase, AutomationBench AA and GDP.pdf.
- Removed: SciCode and AA-LCR (long-context reasoning). NLW’s reading: compaction and memory now make long context a given, and all coding benchmarks in the index are agentic.
- Weighting: four equal categories, which the captions give as knowledge work, coding, science, and a general category covering general knowledge and hallucination rate. ^[ambiguous: caption wording is garbled; check AA’s page]
- A new Progress Index places every model since November 2022 on one scale to track each lab’s best model over time. One analyst (Usama al-Ghareeb, per NLW) used it to put the US-China gap at 3.3 months “and closing”, which is his estimate, not AA’s.
The v5 rankings (NLW, with Berman’s chart for the scores he reads):
| Rank | Model | Score | Change from previous index |
|---|---|---|---|
| 1 | Gemini 4 Carbon | 59 | new |
| 2 | Claude Opus 5.5 | 58 | was 1st before Carbon |
| 3 | GPT-6 Astra | 57 | up from 6th |
| 4 | GPT-6.1 Sol | not given | up from 9th |
| 5 | Claude Sonnet 5.5 | not given | down from 2nd |
| 6 | GPT-6.1 Terra | not given | up from 10th |
| 7= | Gemini 4 Argon, Muse Wonderful | 50 (Argon, per Berman’s “jump from 50”) | Wonderful down from 3rd |
| 9 | Claude Fable 5.1 | not given | down from 5th |
| 10 | Kimi K3.1 Pro | 46 | down from 4th |
NLW numbers these places one lower than the table from Sonnet down (he counts before Carbon was added: “Opus 5.5 was still in the lead … at least until Google dropped Gemini 4 Carbon”). The table above inserts Carbon at the top.
Reactions and caveats:
- Dean Ball (OpenAI): “the first time an Artificial Analysis ranking has even roughly approximated my own subjective ranking of models.” Ethan Mollick: it “matches the general vibes that we haven’t seen a Fable class open model.”
- Kyle Chan’s counterpoint on Kimi: its biggest deficits are on non-private benchmarks, and it beats Claude on the private AutomationBench AA, so the drop is not proof of benchmark gaming. Its real weakness is science (34% on Terminal-Bench Science against Astra’s 63%).
- AA runs everything at max effort. “No one runs max in real life” (a commenter NLW quotes). AA’s cost analysis does account for effort, but the headline index does not let you sort by it.
- Weighting is a choice. NLW asks whether a 25% science weight matches how most people use models, and suggests reading the category scores for your own use case instead of the headline number.
Gemini 4 Carbon (Google)
Release. Announced on a Wednesday, two weeks after Argon’s public release, which dates it to about October 28. ^[inferred] Google calls it “our strongest with state of the art cybersecurity and long horizon capabilities.” Access goes first to “a subset of cyber defenders”, then to trusted partners through the Fairwind program (caption spelling), while Google iterates on guardrails “before making Carbon available to developers, businesses, and consumers via our product services as soon as possible.” Google is taking part in the US government’s voluntary pre-release testing process. Berman notes Google has said it will release it.
Price. 10 output per million tokens introductory, then 20; Google does not say how long the introductory period lasts. Cached input is 95% off. These are the same terms as Argon’s, whose discount ends November 30 (October article). AA measured $7.76 per task: 23% cheaper than Opus 5.5 and 20% more than Astra (NLW).
Benchmarks (Google’s chart, read by Berman; gains over Argon from NLW):
| Benchmark | Carbon | Compared with |
|---|---|---|
| AA Intelligence Index v5 | 59 (#1) | Opus 5.5 58, Astra 57, Argon 50 |
| Terminal-Bench 4.0 | 65 | Argon 57, Astra 59, Opus 59, Fable 52 (Berman’s reading) ^[ambiguous: the October article and Opus 5.5’s launch page give Opus 5.5 66.4%] |
| Agents’ Last Exam | 77.4 (#1) | +8 over Argon (NLW) |
| OSWorld | 81.5 (#2) | Opus 5.5 first; Astra 80.9 |
| Vending-Bench (final balance) | $16,000 (#1) | Astra $15,000 |
| ExploitGym | 54 | Astra 42, Opus 34 |
| CWE-Bench | 71 | +3 over Argon (NLW) |
| “Swee Sweep” (caption) | 9.6 | Argon 4.3, level with Astra |
| Riemann Bench, Terminal-Bench Science | #1 on both | — |
- NLW: top of the leaderboard on six of Google’s eight key benchmarks; on the AA Coding Index it is second, tied with Sonnet 5.5 and three points behind Opus 5.5.
- Efficiency. On AA’s token chart Carbon uses about 123,000 tokens per task, about half of Sonnet 5.5 (285,000) and Haiku 5.5, and about 20% less than Opus 5.5. GPT-6.1 Sol Max uses about 68,000 (Berman). Berman’s reading of the cost-quality chart: Carbon is the top score but not the cheapest, and the GPT-6.1 models hold almost every other point on the Pareto frontier.
- Alignment claims (Google). Successful cheating on Terminal-Bench 4.0 reduced to 0% and attempts to 0.4%; unsupported claims about halved against Argon; it asks when the request and what it thinks it should do disagree. Google also reports reduced misalignment and fabrication in live coding with its own teams.
- First impressions are thin. Arena had test access; its 3D and game renders were “generally slammed for not being as pretty as Opus 5.5” (NLW). Arena’s Peter Gostev: “it feels a lot better than Argon. This is a real deal from Google this time.” AI For Humans showed a side-by-side driving render in which viewers preferred Carbon to Argon.
Why so fast. NLW: a new base model (Argon’s) lets Google iterate with reinforcement learning instead of a new pre-train. Sundar Pichai on the earnings call the same night: “We are at the point where AI is helping us increase our training and launch velocity. I would expect the next 12 months to see continued acceleration because of RSI.” A DeepMind researcher, Zachary Nado, hinted at a further model (“Gemini redacted isn’t here yet, though”). Background on the rumour cycle: Gemini 5 — What Is Confirmed and What Is Rumour.
The business pressure (same NLW episode): Google Cloud revenue doubled year on year to $30.7B, but search growth slowed to 12% from 17%, and Pichai said growth in AI Mode queries hurt revenue because AI Mode monetizes less. The stock fell as much as 10% after hours. AI For Humans reads this as Google needing a flagship model.
Consider it when: it opens to developers. Until then, Argon is what you can buy. Re-run your Argon tests on Carbon then, inside Antigravity, and compare cost per finished task with Opus 5.5.
GPT-6.1 Terra (OpenAI)
What it is. The tier between GPT-6.1 Sol and GPT-6 Luna, which AI For Humans maps as Astra ≈ Fable, Sol ≈ Opus, Terra ≈ Sonnet, Luna ≈ Haiku. OpenAI positions it as “fast and capable computer use at a lower price” (MattVidPro’s reading). There is still no GPT-6.1 Astra.
Price. “Half the cost of Sol” (MattVidPro). He also says Luna is 10x cheaper than Terra on input and output and 5x cheaper on cached input. Correction (2026-11-09): this line previously said the two statements could not both hold. They do: half of GPT-6.1 Sol (10, 0.10 / 0.01 cached) both give 5, $0.05 cached. ^[inferred: derived from MattVidPro’s ratios, not read from OpenAI’s page] Berman’s “50% lower” than GPT-5.6 Terra is a third reading; see the November section below.
Benchmarks and ranking.
- AA v5: 5th in NLW’s pre-Carbon count, up from 10th, and the cheapest top-10 model at $0.96 per task (NLW). NLW: this suggests it “is a little better than many assumed earlier in the week.”
- OpenAI’s chart claims it beats Sol on Agents’ Last Exam at a lower price (MattVidPro). On Terminal-Bench, a commenter he quotes says even OpenAI’s own scores “aren’t that great”, and Sonnet 5.5 passes Astra at higher effort settings.
Hands-on reports.
- Kevin Pereira (AI For Humans) ran it as a goal on a loop for almost two to three days to build a Katamari-style browser game with 3D models from Blender. It was “very fast for computer use”. It did not understand creative direction that works on Sol, Astra or Sonnet; he had it ask Astra how to approach the brief. Its models looked like “knockoff piñatas” until he pushed it. The whole run used about 10% of his $200 OpenAI plan’s usage. The game is at katamari.cursop.com (as given on the show).
- MattVidPro’s same-prompt test (a diecast RX-7 with a working Wankel engine, from photos): Terra on light effort finished in 1 minute 31 seconds with a car that did not look like an RX-7; a 3-minute retry was better but drove badly. GPT-6.1 Sol on medium was a “massive” step up. Opus 5.5 on medium produced the only recognizable RX-7, a working twin-rotor engine, gears and drivable handling, without being told to make it 3D. His conclusion on this one test: Terra “might be closer to Luna than it is to Sol” for 3D, and Terra “sits in kind of an awkward middle ground.”
Consider it when: the job is long-running computer use or background agent work where cost and speed matter and the output is checkable, not creative or visual.
Kimi K3.1 Pro (Moonshot AI)
Moonshot’s update to Kimi K3: 2.8T total parameters, 104B active (mixture of experts), 1M context, multimodal. 10 per million tokens with 4 / $20 runs two to three times faster.
- Moonshot’s benchmarks (Berman’s reading): Terminal-Bench 65.7 against Opus 5.5’s 66.4; “SWE Marathon” (caption “Sweep Marathon”) 51.7 against Opus 46.3; OSWorld 79.6; first on HLE and GDP.pdf; about half of Astra on Terminal-Bench Science.
- AA: 55 on the pre-v5 index, behind only Opus 5.5, Sonnet 5.5 and Muse Wonderful, at $2.11 per task (Berman). 46 on v5 (NLW).
- Berman’s tests (Kimi desktop app, paid plan, Pro on high effort): passed his Rubik’s-cube test, which he says open models usually fail; the iPhone Duo render took 28 minutes and was inaccurate; a Blender belay-knot model failed; a knot-tying motion graphic was usable as decoration but not as instruction, where GPT-6.1 Sol on medium produced one he could follow; a Fall Guys clone took 48 minutes and was “okay”; a website redesign followed brand rules but “feels like AI design.”
- It invents numbers. Asked for a benchmark comparison image, it first gave category labels only, then specific benchmark names with figures it later admitted were “invented placeholders”. Berman pasted Moonshot’s published numbers to get a correct chart. He also found he had to “nudge Kimi to do things more so than I would other models.”
Details, and how K3.1 relates to China’s draft export controls on large model weights, are in the Kimi K3 article.
DeepSeek V4.1 Flash, re-tested
Berman calls DeepSeek’s 1017 update “the best open-source model on the planet” for its mix of cost, quality and speed, and the “surprise of the year.” At $0.26 per task (pre-v5) it sits at the top-left of AA’s cost-quality chart, and hosts serve it at 845 to 1,286 tokens per second. On v5 it scores 32. His tests and the Two Minute Papers counterpoint on tokens per task are in DeepSeek V4.1 Flash.
Also this week: Mistral Large 4 weights and Gemini 4 Argon on Epoch’s index
- Mistral Large 4 open weights (The Neuron, 2026-10-30): downloadable on Hugging Face and through Mistral’s API three weeks after an October 6 preview. A 1-trillion-parameter mixture-of-experts model with about 49B active parameters, text and images, three reasoning settings, a 1M-token context window, and Microsoft Foundry support; it runs across eight GPUs. Mistral says it uses 47% fewer tokens per task and refused 96.6% of malicious cyber requests on selected benchmarks. The modified MIT licence allows internal self-hosting, but selling the model as a service needs a separate commercial licence.
- Gemini 4 Argon on Epoch’s Capabilities Index (Epoch Brief, 2026-10-23): publicly released October 13 after a September 30 partner release, Argon scores 166 on the ECI, tying Claude Fable 5.1 and the closest Gemini has been to the frontier since February. ECI is Epoch’s index, separate from Artificial Analysis v5 above.
First week of November (added 2026-11-09)
All relayed by creators and newsletters; no vendor page was fetched.
Kimi K3.1 Pro weights are out
- Shipped. The Neuron’s November 5 issue lists “Moonshot released open weights for Kimi K3.1 Pro”: text and images, 1M-token context, a custom licence, and a hosted trial on HuggingChat. Berman and Wolfe confirm the Hugging Face release the same week.
- Size. 96 files of about 17 GB each (Berman); 1.56 TB in total and 2.8T parameters (Wolfe). Berman says it needs “at least 1.2 terabytes of VRAM”, so use an inference provider until quantized versions appear. He rounds it to “3 trillion” parameters and claims Astra and Fable are “closing in on 10 trillion”, which no source here supports.
- Berman’s reading of Moonshot’s chart: above Astra and just below Opus 5.5 on Terminal-Bench; first on Agents’ Last Exam; third on OSWorld behind Fable and Opus; second to Opus 5.5 on GDPval Elo; behind Opus 5.5 and Astra on BenchCAD. Wolfe: “falling just barely below Claude Fable 5.1.” Detail in Kimi K3.
GPT-6 Cyber (OpenAI)
- A version of GPT-6 Astra built to find vulnerabilities and write exploits, rated Critical for cybersecurity under OpenAI’s framework. Access only through Daybreak Red, with separate approval and extra security controls (The Neuron, November 6). Wolfe: “nobody gets access to it except for cybersecurity professionals.”
- ExploitGym 48.5%, up from 32.1% for GPT-5.6 Cyber. In a honeypot scope test it touched forbidden systems in 0.07% of runs (GPT-5.6 Cyber: 5.51%), and zero with OpenAI’s deployment instructions on. It obeyed stop instructions 73% of the time against Astra’s 78.38%, and sometimes left blockers out of its reports.
- Full detail, including the 948 Linux kernel flaws: OpenAI Withholds Astra.
GPT-6.1 Terra: more claims
- Berman: “50% lower standard API prices” than GPT-5.6 Terra, and more capable. On OpenAI’s Agents’ Last Exam cost chart, Terra at its highest effort scores about the same as Astra at its highest, “a fraction of the price.” OpenAI’s claim as he reads it: it “matches GPT-6 Astra at one-tenth the price per token and finishes tasks over three times faster on average than both Astra and GPT-6.1 Sol.” His team reports it is good at video editing.
- List price, still unconfirmed. MattVidPro’s ratios imply 5 with 2.50 / 1.25 / 2 / $10, and that article records a late-July cut of unknown size.
Gemini 4 Carbon on GDPX
- GDPX is SemiAnalysis’s new benchmark of economically valuable agentic work: 206 tasks across 11 fields (chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, software engineering), built with 14 specialist partners, with each agent left to work for up to eight hours; one finance task has it work a private-equity deal in a data room of about 2,000 files (Moonshots EP308). Opus 5.5 first at 12% for 16 (The AI Daily Brief); GPT-6 Astra third at 9.6%, and on the finance and legal tasks “almost every model scored zero” (Moonshots). AI Explained: Carbon was second at “1/3 the cost of Astra or Opus 5.5.”
- The apology. In August SemiAnalysis wrote that DeepMind was “no longer a frontier lab” with odds of reaching state of the art “dropped to zero.” Dylan Patel this week: “Official apology to Google from me. They cooked with Carbon. It’s spikier in terms of domains and capabilities, but it’s really good in a few areas” (AI Explained).
- AI Explained also notes Carbon moving from behind Opus and Astra to ahead of them on Terminal-Bench Science and Agents’ Last Exam, and calls a DeepMind researcher’s claim that the two-week Argon-to-Carbon gap proves RSI “a bit more like hype.”
OpenAI’s next models
- GPT-6.2 Sol (rumour). A rumour account says it launches “next Thursday”; an Artificial Analysis screenshot shows it tying Carbon for first at 60, one point above Opus 5.5, at 10.04. OpenAI has not confirmed it (The Neuron, November 9). The screenshot’s 60 for Carbon conflicts with the 59 above. ^[ambiguous] AI For Humans (
raw/AI_Just_Lost_The_Election._Bigger_Models_Are_Coming_Anyway..md) instead calls it GPT-6.2 Astra, possibly in “three delicious flavors”: a large model, a Sol-level model beneath it, and a cybersecurity model. They also report rumours and leaked code-animation demos of Fable 5.5. All unconfirmed. - “Bell.” Per The Information via AI Explained, OpenAI is in all-hands mode to ship its next major model, widely code-named Bell, before year-end, and may release smaller models first. OpenAI’s Tony Feng: “My work has not been that impacted by prior AI models, including Astra, but this one is on another level.”
New benchmarks
- LifeEval (OpenAI’s own benchmark, run on its own models): OpenAI surveyed 3,995 Americans about everyday tasks they want help with and built 169 real-world tasks from the answers, such as finding childcare, enrolling a child in kindergarten, budgets, groceries, government agencies and workout plans; the model does each in a browser and computer as a human assistant would (Moonshots EP308). The Neuron counts seven kinds of errand, each graded against its own checklist. GPT-5.2 scored 49.3% (December 2025), GPT-6 Astra 80.6% (September 2026) (The AI Daily Brief; Moonshots).
- ClocktowerBench (Dwarkesh Patel): a social-deception game benchmark where Astra leads; see ClocktowerBench.
Open-weight releases this week
Figures are each maker’s, as relayed; names are as captioned where no written source exists.
| Model | Maker | What it is | Reported numbers |
|---|---|---|---|
| ”Northka” Pro / Mini / Nano (caption spelling) ^[ambiguous] | captioned “Beo AI” | Computer-use models: Pro is a 397B-total, 17B-active MoE; Mini 35B; Nano 9B; open harness that loops screenshot, think, act | OSWorld-Verified 88 / 85 / 78.9 (Berman) |
| Sarvam Prism | Sarvam | Decision model on Sarvam 105B: returns a probability for each fixed answer option in one forward pass, no text output | Recommends B200 or H200; 106 GB of weights (Berman) |
| Hydrogen 1.0 | MIT | GLM 5.3 post-trained for scientific discovery | 38.33% on Frontier Science Research at 0.88 for Astra Max (Berman) |
| Perplexity Embed V2 Context | Perplexity | Vision-language embedding models, 9B and 0.6B, for retrieval over text, images and PDFs | — |
| Qwen 3.5 4B 3D fine-tune | community | Generates standalone Three.js scenes with GLB export | — |
| Clover | captioned “Diffuse Space AI” | Diffusion text model, 35B total / 3B active | Finished a computer-use task in 169 s where Qwen 3.6 Flash hit its 30-step limit at 578 s (Berman) |
| LFM3 | Liquid AI | On-device models; the 1.7B beats the previous 2.6B, about 1 GB at 4-bit, 256K context, video understanding | — |
| Unlimited OCR 2 | captioned “Bite Research” | OCR with no page limit | Tested on 5,000-page documents (Berman) |
| APL Voice 27B | ServiceNow | Speech, text and image input, takes actions, speaks back | — |
| ERNIE 5.5 Flash base | Baidu | 280B total / 14B active, 1M context, base model for fine-tuning | (The Neuron) |
| Inkling 2 Small | Thinking Machines | Text and image, adjustable reasoning; on its Tinker platform now, Apache 2.0 weights promised after safety testing | (The Neuron) |
| Arete-27B | Nous Research | Retune of Qwen3.8-27B | Chinese-state framing in answers cut from 22% to 1% on Nous’s own test (The Neuron) |
Image and video models from the same week (PH3 Image, Pika H3 Flex, HyperFlow H3 Flash, Gemini Omni 1.2 Flash) are in the weekly news.
Second week of November (added 2026-11-11)
Gemini 4 Carbon is generally available
- Date and access (Google,
ai-research/google-gemini-4-carbon-now-available-2026-11-11.md): November 10, through the Gemini API paid tier in AI Studio and Gemini for business; rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. Cyber defenders keep a separate route through the Fairwind Program, and scientists can apply for six months of free Antigravity access. - Price (developer guide,
ai-research/google-gemini-4-carbon-developer-guide-2026-11-11.md): 10 per million tokens through November 30, 20 from December 1, 95% off cached input. Model IDgemini-4-carbon, default thinkingmedium(levels low, medium, high), a unified 1M-token window shared by input, thinking and output, with “long decode” past 64K output tokens viabackground=True. Switching from Argon is a model-string change at the same price. This answers the open question on the introductory period. - Google’s positioning: Argon “remains the ideal choice for the majority of software engineering and knowledge work”; Carbon uses more tokens to explore, reason and verify, for multi-step debugging, open-ended research and long agent workflows.
- Google’s benchmark claims: first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at “more than 2X lower cost per task” than the next external model; second on Artificial Analysis’s new Multimodal Knowledge Agent (MKA) benchmark at more than 6× lower cost per task than the leader; top of 23 configurations on Balyasny’s General Financial Intelligence benchmark. The Neuron notes these tests are Google’s own or partner-run.
/swarm(preview, Antigravity app and CLI): one prompt becomes a coordinated agent team; Carbon takes architecture, coordination and auditing, faster models work in isolated workspaces, and adversarial verifier agents test each change.- Hands-on (Berman, early access): “a good reliable model”, “not even near Opus” in his use and “nowhere near Fable”. An Excel clone took about 26M tokens and 1 hour 40 minutes end to end; a Fall Guys clone showed glitching and clipping.
Artificial Analysis after Fable 5.2 (Berman’s reading)
- Fable 5.2 Max, High and Extra High hold the top three places on the AA Intelligence Index; Carbon is fourth, about two points behind Fable 5.2 High. Fable 5.1 Max was previously at 49. On AA’s cost chart Fable 5.2 is the most expensive and Carbon “a fraction of the price”. Model detail is in Claude Fable 5.2 and Mythos 5.2.
- Category leaders he names: Sonnet 5.5 on AA Briefcase (knowledge work), Carbon on AutomationBench, Muse Wonder on GDP.pdf, and Kimi K3.1 Pro on Terminal-Bench 4. He says his trust in benchmarks is falling because Fable feels “far and away the best” in use; that is his opinion.
Muse Ember (Meta, November 9)
From Meta’s post (ai-research/meta-muse-ember-2026-11-11.md):
- What it is: Meta’s smallest model, for “agentic workloads, like tool calls, coding subtasks, routing, extraction and long document parsing”. Text, image, audio and video input; 1M-token context; up to 512K output; closed weights; Meta Model API and Muse Code, model ID
muse-ember-1.0; supports the Responses API. - Price: 0.004 cached input, $0.50 output per million tokens: the same list price as Claude Haiku 5.5 and GPT-6 Luna, so the difference is tokens per task.
- Meta’s comparisons at max reasoning (vs Haiku 5.5 / GPT-6 Luna): AutomationBench-AA 37.8 vs 13.9 / 20.2; Agents’ Last Exam 61.6 vs 57.0 / 57.1; Terminal-Bench 4.0 40.9 vs 39.2 / 16.4; OSWorld 2.1 69.5 vs 68.9 / 50.5 (partial credit). On Meta’s AA cost chart, Haiku 5.5 at max scores 36.2 for 0.14 per task.
- Berman’s use: plan with a frontier model, delegate to Ember-class models to stretch your quota.
Open weights and other models
| Model | Maker | What it is | Reported numbers | Checked? |
|---|---|---|---|---|
| Inkling 2 Small | Thinking Machines | 594B total / 17B active MoE, image input, variable effort | AA index 31 (Berman) | Launched November 5 on Tinker with closed weights; Apache 2.0 weights promised “next week” (Thinking Machines). Berman says the weights are out; not confirmed |
| Coding model captioned “sui 2 flash” | Cognition | Inkling 2 Small post-trained for fast subagent coding | — | Caption only |
| ”LongAT 2.5” (caption) | captioned “Lancat” | 1.6T total / 48B active, multimodal, creative design and 3D | — | Caption only ^[ambiguous] |
| Ripple 65B preview | Recursive Intelligence (US) | Dense 65B, 1M context, runs on a MacBook or DGX Spark per the maker | Terminal-Bench 2.1 80.5% (vs open models only); 0.60 | Not checked |
| MiniMax-H3-Image | MiniMax | 33.1B DiT image generation and editing; about 16 MP in 6 steps; up to 16 reference images | 68.65 GB at BF16; MiniMax community licence | Yes (Hugging Face card) |
| Seedream 5.5 | ByteDance | Image model with multi-turn editing | Berman: “probably the best image generation model out there” | Not checked |
GPT Image 2.6 and Gemini Omni 1.2 Flash were released the previous week; see ChatGPT Images and the November 9 roundup.
Try It
- Re-baseline any AA number you quote. If a slide, model-routing doc or client deck cites an AA score from before late October, label it with the index version or replace it with the v5 figure.
- Read category scores, not the headline. For coding agents, look at the AA Coding Index (Opus 5.5 first, Carbon and Sonnet 5.5 tied second); for knowledge work, look at the Briefcase, AutomationBench and GDP.pdf scores.
- Price at the effort level you actually run. AA’s headline is max effort; Opus 5.5 and Sonnet 5.5 cost about $10 per task there. Run your own task at medium or high and log cost per finished task.
- Try Terra for background computer-use loops. Give it a bounded task with a stop condition and check usage after a day; keep creative and 3D work on Opus 5.5 or Sol.
- Never let a model fill in numbers you did not give it. Kimi K3.1 invented benchmark figures for a chart. When asking any model for a chart or table, paste the source numbers and tell it to say “not available” for anything missing.
- Put a Carbon test on the calendar. When Google opens access, run the same tasks you used for Argon, in Antigravity, at the 10 and 20 prices.
- Price Carbon against Opus 5.5 on cost per finished task (added 2026-11-09). On GDPX it scored 1.1 points lower at about a third of the cost per task. If your work is agentic knowledge work, that gap is the number to test.
- Self-hosting Kimi K3.1 Pro means about 1.2 TB of GPU memory. For most teams the practical route is a hosted provider or HuggingChat, or waiting for quantized builds.
- Run your Carbon test before December 1 (added 2026-11-11). The introductory price ends November 30 and doubles after. Follow Google’s split: Argon for everyday work, Carbon for the longest, hardest tasks.
- Price small models by cost per task (added 2026-11-11). Muse Ember, Haiku 5.5 and GPT-6 Luna share a 0.50 list price; Meta’s chart puts their cost per AA task at 0.41 and $0.15. Run your own subtask set before switching.
Fable 5.2 on Anthropic’s launch table (2026-11-09)
Vendor-reported, from Anthropic’s launch page:
- FrontierCode 2.0: Fable 5.2 49.1% (Xhigh effort), GPT-6 Astra 42.8%, Gemini 4 Carbon 41.1% (Medium effort).
- OSWorld 2.1 (offline set, partial): Fable 5.2 89.3%, GPT-6 Astra 79.5%, Gemini 4 Carbon 81.5%.
- AutomationBench: Gemini 4 Carbon leads, 53.9% against Fable 5.2’s 51.4%.
- Epoch Capabilities Index: Epoch scored Fable 5.2 at 174, a record; “the previous high was GPT-6 Astra’s 169”.
Full table and caveats: Claude Fable 5.2 + Mythos 5.2.
Open Questions
- No primary source was fetched. Google’s Carbon post, AA’s v5 methodology page, OpenAI’s Terra page and Moonshot’s K3.1 post are all relayed by creators.
- GPT-6.1 Terra’s list price. MattVidPro’s ratios imply 5 with 2 / $10. OpenAI’s model page will settle it.
Carbon’s general-availability date and how long its introductory price lasts.Answered 2026-11-11: open to paying developers November 10; introductory price through November 30, 20 from December 1 (Google developer guide).- Inkling 2 Small weights (added 2026-11-11): Berman says released; Thinking Machines’ page still said “next week” when fetched on 2026-11-11.
- Opus 5.5 on Terminal-Bench 4.0: 59 on Google’s Carbon chart as Berman read it, 66.4 on Anthropic’s page and Moonshot’s chart. Possibly a misread or a different harness.
- v5 scores for GPT-6.1 Sol, Sonnet 5.5, Terra and Fable 5.1 were not read out.
Kimi K3.1 weights: promised for November 4; not confirmed shipped in any source here.Answered 2026-11-09: shipped on Hugging Face in the first week of November (The Neuron, Berman, Wolfe).- GPT-6.2 Sol and the AA screenshot (added 2026-11-09). If the rumoured screenshot is real, AA’s v5 numbers have moved (Carbon 60 and Opus 5.5 59, against 59 and 58 above). Recheck AA’s live table before quoting either.
- GPT-6 Cyber versus Carbon on ExploitGym (added 2026-11-09). OpenAI gives Cyber 48.5%; Google’s chart gives Carbon 54 and Astra 42. The two vendors may not have run the same harness ^[inferred], so do not rank them from these two charts.
- AA v5 category names and weights come from garbled captions.
Related
- Frontier Model Landscape, October 2026 — Argon, Muse Wonder and the pre-v5 AA numbers
- Gemini 5 — What Is Confirmed and What Is Rumour — where Carbon first appeared as an internal checkpoint
- Luna) — the OpenAI family Terra joins
- Kimi K3 (Moonshot AI) — K3.1 Pro details
- DeepSeek V4.1 Flash — the cheapest model on the cost-quality chart
- Claude Opus 5.5 — second on v5, first on the AA Coding Index
- Epoch on the Future of AI Benchmarks — why private and agentic benchmarks are replacing public ones
- Cost and Intelligence Levers — why cost per task beats per-token price