Skip to content
All writing

Which AI model is best for coding agents in November 2026?

Aarav Pundir · FounderAI models for coding agents6 min

Claude Fable 5.2, GPT-6.2 Astra and Gemini 4 all shipped in November 2026. Their prices, independent scores, and what each costs per finished task.

There is no single best model for coding agents in November 2026. Three labs released new models within ten days, and the top of every independent test is close. Anthropic's Claude Fable 5.2 leads the Artificial Analysis index at 65 and the main web development arena. OpenAI's GPT-6.2 Astra ties it on Epoch AI's capabilities index and costs far less per task. Google's Gemini 4 family spans the range from cheap to frontier. On Terminal-Bench 4.1, the closest public test to agent work, the top five results sit within about the test's own margin of error. Choose by what a finished task costs you, and let the agent's harness and effort setting count as much as the model.

What did Anthropic release in November 2026?

Claude Fable 5.2 and Claude Mythos 5.2, on 9 November. They are one model sold in two tiers. Fable is open to every customer and carries safeguards for cyber and biology work. Mythos goes only to organizations in Anthropic's cyber and life sciences verification programs.

The 5.2 release writes output up to 30% faster than 5.1 did, and it costs less. Anthropic's price page lists $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Both have a one million token context window and 128,000 tokens of output. On Anthropic's own coding test, FrontierSWE v2, Fable 5.2 scored 69.9%.

The rest of the Claude line was refreshed in the autumn:

  • Claude Opus 5.5, 22 September, at $4 and $20.
  • Claude Sonnet 5.5, 28 September, at $2 and $10.
  • Claude Haiku 5.5, 7 October, at $0.10 and $0.50 for prompts under 100,000 tokens.

Anthropic's own guidance is to start with Opus 5.5 for most work and keep Fable 5.2 for the most demanding long tasks.

What did OpenAI release?

GPT-6.2 Astra and GPT-6.2 Sol, on 17 November, in ChatGPT, Codex and the API. Only Astra got cheaper. Its API price is $8 per million input tokens and $40 per million output tokens for prompts up to 272,000 tokens, down from $10 and $50 for GPT-6 Astra. Sol stays at $2 and $10. Astra takes about one million tokens of context.

OpenAI's headline claim is a score of 78.3% on a test called Agents' Last Exam, against 73.1% for GPT-6 Astra, as VentureBeat read OpenAI's chart. Treat that number with care. On 18 October Epoch AI rated that benchmark "Flawed": 22 of the 50 tasks it sampled had grading errors that let an agent score full marks without doing the work. The launch also had a poor reception from some heavy users, which is in VentureBeat's headline.

What did Google release?

Four Gemini 4 models, rolled out over seven weeks.

ModelReleasedPrice per million tokensFor
Gemini 4 Argon30 September$2 and $10, then $4 and $20Cyber defenders first
Gemini 4 Carbon10 November$2 and $10 until 30 November, then $4 and $20Google's most capable
Gemini 4 Hydrogen12 November$1.50 and $7.50Fast, everyday agent work
Gemini 4 Neon18 November$3 and $15Mid-tier coding

Google claims Neon is first on the Artificial Analysis coding agent index, at 58.6. Note that Carbon's introductory price ends on 30 November, when it doubles.

Which model scores best on independent tests?

It depends on the test, and the leaders are close on all of them.

Artificial Analysis Intelligence Index, a broad mix of tests, read on 19 November:

Model and effortScoreCost per index task
Claude Fable 5.2, max65$26.54
GPT-6.2 Astra, max62$4.00
Gemini 4 Neon, xhigh61$6.44
Gemini 4 Carbon, high60$7.76
Claude Opus 5.5, max59$10.04
GPT-6.2 Sol, max57$1.01

Terminal-Bench 4.1 measures an agent doing real tasks in a terminal, so it tests the agent program as well as the model. GitHub Copilot's own agent leads at 71.2%. Claude Code with Opus 5.5 and Claude Code with Fable 5.2 both score 65.8%. Gemini 4 Carbon in Google's Antigravity command line tool scores 64.5%, and Codex with GPT-6.2 Astra scores 62.1%. The margins are about three points, so the top five are roughly tied.

Epoch AI puts Fable 5.2 and GPT-6.2 Astra level at 174 on its capabilities index. On its own hand-graded work tasks, Astra averaged 17 points below Fable 5.2. No model met Epoch's standard for a member of its staff.

Arena, where people vote on answers, has Fable 5.2 first in web development at 1824, just ahead of Opus 5.5 at 1818 and Astra at 1797.

Why does Fable 5.2 cost so much more per task?

Because at maximum effort it thinks for much longer. Fable 5.2 and GPT-6.2 Astra have the same list price, yet Fable cost 6.6 times as much per task on the index. At maximum effort it used 275 million output tokens across the index, against a median of 65.1 million, and took 41.8 minutes per task. At high effort it scored 62, the same as Astra at maximum, in 15.1 minutes.

So the effort setting is a pricing decision. The same model can cost several times more for a few points of score, and the right setting differs by task.

Did list prices go down?

At the top, yes, by about a fifth. Fable went from $10 and $50 to $8 and $40. GPT-6 Astra to GPT-6.2 Astra made the same cut. Opus 5 to Opus 5.5 went from $5 and $25 to $4 and $20. Gemini 4 Carbon goes the other way on 1 December, when its introductory price ends.

Why can you no longer measure how long an agent can work?

Because the best known measure stopped keeping up. METR's time horizon, the length of task a model can finish half the time, became a standard way to track agent progress. METR's page says it is no longer actively updated. The last model it added, an early Claude Mythos preview in May, was estimated at 16 hours or more, and METR says it cannot reliably measure above 16 hours with its current tasks. There is no METR figure for any model released this month.

What happened to Claude Fable in June?

It was switched off for everyone for three weeks. Fable 5 and Mythos 5 launched on 9 June. On 12 June a US export control directive barred access by foreign nationals, and because Anthropic could not check nationality in real time it disabled both models for all customers. The controls were lifted on 30 June, and Fable 5 came back worldwide on 1 July with a new safety classifier, as Anthropic described when it redeployed it.

For three weeks, work built on those two models could not run, for reasons no customer controlled. That is the strongest argument for keeping a second model ready.

How should a team choose a model for its agents?

By measuring finished work on its own tasks.

  • Price a finished task, not a token. Run the same few real tasks on two or three models and compare the bill for the ones that passed.
  • Test the agent program too. Terminal-Bench shows the harness changing results by several points with the same model.
  • Set effort by task. High effort is often most of the score for a fraction of the cost.
  • Keep the model a setting. June showed that access can stop overnight.
  • Judge every model's work against the same test. A written definition of done is what makes a switch safe.

Cognibl's cloud agents run Anthropic's models through Claude Code and OpenAI's through Codex. You choose the provider, the model and the reasoning level on each agent's settings, and can change them at any time. Every turn records the model its tokens were spent on, and every task is checked against the same definition of done whichever model did the work. See how cloud agents choose their model.