The four highest-scoring models on the BenchLM leaderboard all launched or opened to the public after November 8.
Claude Fable 5.2 leads at 93.2, ahead of GPT-6.2 Astra at 91.1, Gemini 4 Carbon at 88.1, and Gemini 4 Neon at 86.3. Two of those ranks are days old, and Neon's rests on four independent evaluators.
This issue covers November 8 to 21, 2026, since we skipped November 14. Scores and ranks come from our data on November 19. Prices are per million input and output tokens.
A launch-week rank has fewer outside tests under it
Our overall score takes two kinds of results. Some come from the lab that built the model, and some come from independent evaluators who ran their own test. Lab-run results count for less. Independent ones take longer, because a benchmark's maintainer needs access to the model, then has to run the suite and publish the row.
So a model's rank on launch day leans on the few boards that move fast, and it fills in over the following weeks. You can see the fill in the top seven.
Fable 5.2 came out on November 9, and ten days later 19 independent evaluators had tested it. GPT-6.2 Astra launched on November 17 and had 11 by November 19. Neon launched on November 18 and had four the next day, with one result from each. GPT-6 Astra, out since September 3, has 37.
Thin doesn't mean inflated. Neon's four outside results are strong enough to put it third on our agentic board at 91.6. They're just few, so its fourth place could settle higher or lower as more boards report.
A new model's rank is only as old as its evidence. Before you move work to a model that launched this week, count how many of its results someone other than the lab ran. Each model page labels who ran every score.
GPT-6.2 Astra and Claude Fable 5.2 now cost the same
Anthropic released Fable 5.2 on November 9 at $8 input and $40 output, down from $10 and $50 for Fable 5.1. OpenAI launched GPT-6.2 Astra on November 17 at the same $8 and $40, which is 20% below GPT-6 Astra. Price no longer separates the two flagships, so the job has to.
On our category boards they're level on coding, at 92.3 each, and within half a point on agentic work. Fable 5.2's clear lead is reasoning, 93.7 to 87.5. Outside tests published this fortnight lean the other way on long engineering tasks. On Bug Hunt Bench, GPT-6.2 Astra fixed an average of 62.7 of 105 planted bugs, against 58.0 for Fable 5.2. On Mercor's OpenWorldEngineer-Bench, it leads 67.2 to 54.6, though each Astra attempt cost $190.66 against $105.08 for Fable 5.2.
GPT-6.2 Sol is the quieter launch. It replaces GPT-6.1 Sol at the same $2 and $10 and scores 82.6, up from 80.7. That puts it a tenth of a point behind Claude Opus 5.5, which costs $4 and $20. A gap that small is a tie, and Sol has 10 independent evaluators to Opus 5.5's 33, so we'd test both on your own tasks before switching to save money.
OpenAI's system card also reports GPQA Diamond at 97% for Astra and 96% for Sol. OpenAI ran those tests on its own models. It shipped no GPT-6.2 Luna, and GPT-6 Luna stays its efficient model.
Gemini 4 Carbon doubles in price on December 1
Google made Carbon generally available on the paid Gemini API on November 10, after a limited release announced on October 28. It lists at $2 input and $10 output through November 30, then $4 and $20. Gemini 4 Argon follows the same schedule.
Google shipped two more models around it. Gemini 4 Hydrogen arrived on November 12 at $1.50 and $7.50, with a free tier, and ranks 11th at 78.8 on 15 independent evaluators. Neon arrived on November 18 at $3 and $15, on the paid API only.
Take a workload of 100 million input and 20 million output tokens a month. It costs $400 on Carbon today and $800 from December 1. Neon runs the same volume for $600, and Fable 5.2 or GPT-6.2 Astra for $1,600. So in ten days Neon becomes the cheaper of Google's two top models, on a score that four evaluators support against Carbon's 22.
If you run Carbon or Argon in production, redo December's budget now. The Google API pricing page has both dates and the billing tiers.
Fable 5.2 leads overall and trails on five of nine specialist boards
On November 19 we moved the leaderboard to BenchAlign v6.0 and added nine specialist leaderboards. The overall score blends six categories, with agentic work and coding at 25% each. A specialist board asks a narrower question, and the answer often differs.
| Board | Leader | Next |
|---|---|---|
| Cybersecurity | Gemini 4 Carbon, 89.8 | Claude Fable 5.2, 86.7 |
| Documents | Claude Fable 5.2, 81.4 | Gemini 4 Carbon, 80.4 |
| Factuality | Gemini 4 Carbon, 78.7 | GPT-6 Astra, 77.7 |
| Finance | Gemini 4 Carbon, 89.0 | Gemini 4 Argon, 88.7 |
| Frontend | Claude Fable 5.2, 89.5 | Claude Opus 5.5, 82.8 |
| Healthcare | Claude Fable 5.2, 87.3 | Gemini 4 Carbon, 85.5 |
| Legal | Claude Fable 5.2, 86.0 | Claude Opus 5.5, 85.8 |
| Science | GPT-6.2 Astra, 91.1 | Claude Fable 5.2, 90.8 |
| Search | GPT-6.2 Astra, 88.3 | Claude Fable 5.2, 86.1 |
Fable 5.2 leads four boards, Carbon three, and GPT-6.2 Astra two. Read the margins before the names. Finance, legal, and science are each decided by less than half a point, and most rows on these boards carry our Estimated label, which means the rank is visible and the uncertainty is wider. Frontend has the widest gap, at 6.7 points.
One warning about old numbers. The move to v6.0 rescored every model, so scores in this issue don't compare with the November 7 issue. Kimi K3.1 Pro read 75.03 and 12th then. It reads 70.6 and 21st now, and it's still the highest-ranked open-weight model we track.
Most smaller launches came with lab-run numbers only
Microsoft launched MAI Code 2 and MAI Code 2 Flash on November 17, inside GitHub Copilot only. Copilot bills MAI Code 2 at $0.50 and $2.50 until May 17, 2027, then $1 and $5. Microsoft reports 73.2% on DeepSWE 1.1, from its own runs on its own harness, so neither model has a rank.
Tencent's Hy4 flash launched a day earlier at $0.07 and $0.281, billed at half that until November 30. Baidu's ERNIE 5.5 Flash Preview arrived on November 11 with a free OpenRouter route, and Baidu says open weights come by November 25. Both labs ran every launch benchmark themselves. Qwen4 Max is the exception. It went generally available on November 16 at $1.50 and $4.50, and four independent evaluators place it 17th at 74.5.
We also added Elorian's Glimpse, a visual reasoning model at $0.60 and $3, and Liquid AI's d1.2 decision models. OVHcloud announced Olala mini 7A1B on November 17, with 7 billion total and 1 billion active parameters. It promises Apache 2.0 weights, but no repository was public when we checked that day, and it published no benchmarks.
Two new benchmark pages are worth opening. EyeBench asks 100 visual puzzles: GPT-6.2 Astra answered 99, GPT-6.2 Sol 98, and Fable 5.2 86, and the Sol run cost $5.02 against $65.56 for Fable 5.2. SWT-bench Verified scores agents on writing tests that reproduce real GitHub issues.
Anthropic's plans changed too. Max is now one $100 plan, and the old $200 tier is called Ultra. We updated Claude Pro vs Max and priced Claude Cowork on the new lineup.
Three things to watch before the next issue. Baidu's ERNIE weights are due by November 25. November 30 is the last day of $2 and $10 on Carbon and Argon, and of half-price Hy4 flash. And Neon's evaluator count should climb: Fable 5.2 went from launch to 19 in ten days, so fourth place gets its real test soon.
Release dates and sources for every model above are on model updates.
Frequently asked questions
01What was the biggest LLM news in mid-November 2026?
Three labs shipped flagship models in ten days. Anthropic released Claude Fable 5.2 on November 9, OpenAI launched GPT-6.2 Astra and Sol on November 17, and Google released Gemini 4 Neon on November 18. With Gemini 4 Carbon, they hold the top four overall ranks in our November 19 data.
02How much do GPT-6.2 Astra and GPT-6.2 Sol cost?
GPT-6.2 Astra lists at $8 input and $40 output per million tokens, 20% below GPT-6 Astra and the same as Claude Fable 5.2. GPT-6.2 Sol lists at $2 and $10, unchanged from GPT-6.1 Sol. OpenAI shipped no GPT-6.2 Luna and keeps GPT-6 Luna as its efficient model.
03When does Gemini 4 Carbon's API price go up?
December 1, 2026. Google lists Gemini 4 Carbon and Gemini 4 Argon at $2 input and $10 output per million tokens through November 30, then $4 and $20. Gemini 4 Neon, released November 18, lists at $3 and $15, and Gemini 4 Hydrogen at $1.50 and $7.50.
