the best ai model for coding in september 2026, by job

the best ai model for coding as of september 2026 is claude opus 5.5 running in claude code, for anything long and agentic. anthropic's table puts it at 66.4% on terminal-bench 4.0 against 57.9% for gpt-6 astra, at $4 in and $20 out per million tokens. for cheap background agents, gemini 3.8 flash is the best value. for local work, qwen3 coder next is the pick.

that is the short answer. the long one matters because three of the vendors below shipped a new model this week.

what shipped this month

four vendors shipped coding models in september 2026. in order:

• september 1: anthropic released claude fable 5.1, at $10 in and $50 out.
• september 2: google released gemini 3.8 flash, priced at $0.75 in and $3.75 out through december 31.
• september 3: openai shipped gpt-6 astra, its flagship, at $10 in and $50 out with a 1,050,000 token context window.
• september 21: xai, which now publishes as spacexai, released grok 4.7 and made it available in its grok build cli.
• september 22: anthropic released claude opus 5.5, and openai followed the same day with gpt-6 sol and gpt-6 luna.

so any ranking older than a week is already stale. this one will be too by christmas.

which benchmarks still mean anything?

swe-bench verified used to be the scoreboard. it has aged out. none of the september launch posts we read report it. they report terminal-bench 4.0, deepswe v1.1, frontiercode and cursorbench instead.

terminal-bench 4.0 is the most useful of these for agent work. it measures whether a model can finish complex, multi step tasks in a command line, which is what claude code, codex and gemini cli actually do. the older terminal-bench 2.1 is saturated: gemini 3.8 flash scores 89.4% on it and only 19.1% on version 4.0, per its deepmind model card.

one caution. most numbers below are vendor reported. harnesses, effort levels and scaffolds differ, so compare within a column, loosely. the public terminal-bench 4.0 leaderboard does not list opus 5.5 yet.

how do the models compare on price and benchmarks?

sources: prices from claude pricing, the openai model docs, gemini api pricing and xai's pricing table.

terminal-bench 4.0: opus 5.5, fable 5.1 and gpt-6 astra from anthropic's opus 5.5 table, with astra's score as reported by openai; grok 4.7, gemini 3.8 flash and sonnet 5 match the public leaderboard, where astra's best run is 58.2% and fable 5.1's is 57.9%.

deepswe v1.1: astra and fable 5.1 (67.4%) from openai's astra post, fable 5.1 (70.0%) and grok 4.7 (high effort) from xai's grok 4.7 post, gpt-6 sol (max effort) from openai's sol and luna post, gemini 3.8 flash and sonnet 5 from google's model card.

the gap that jumps out: in anthropic's table, opus 5.5 beats gpt-6 astra by 8.5 points on terminal-bench 4.0 while costing 60% less per token. anthropic also says that at default settings it costs 40% less than opus 5 on typical workloads.

claude code vs codex vs gemini cli: does the harness matter?

more than most rankings admit. the model is half the product. the other half is the loop around it: how it plans, reads files, runs tests and recovers.

claude code got opus 5.5 on day one. github, one of anthropic's early testers, says that in vs code opus 5.5 solved more terminal tasks than opus 5 in less than half the steps. fewer steps means fewer tokens, which matters more than list price on a long job.

codex added gpt-6 sol and luna on september 22, rolling out to plus, pro, business, enterprise and edu users, alongside astra. sol at $2 and $10 is openai's value pick, and it matches sonnet 5's list price exactly.

gemini cli is google's terminal agent, and flash is where google competes. at $0.75 input, gemini 3.8 flash posts 73.7% on deepswe v1.1, within half a point of astra's 74.1% at 7.5% of astra's input price. it keeps that intro price until december 31, 2026, then doubles to $1.50 and $7.50.

grok build is xai's own terminal agent, and grok 4.7 is available in it. the cheaper grok build 0.1 model lists at $1 in and $2 out.

what is the best open weight model for coding?

two stand out, for different reasons.

glm-5.3 from z.ai is the strongest open coder we could verify. it is the only open weight model on the terminal-bench 4.0 leaderboard, at 41.8% in claude code, and its card reports 66.9% on deepswe v1.1. it reuses the base model of glm-5.2, which has about 753b total parameters and a 1m token context window.

glm-5.2 ships under mit. glm-5.3 uses its own permissive license, which adds a security review only for model hosting businesses with more than $10 billion in revenue. it is a datacenter model. you will not run it on a laptop.

qwen3 coder next is the local coding model to try. it has 80b total parameters but activates only 3b per token, ships under apache 2.0, and reports 70.6% on swe-bench verified and 44.3% on swe-bench pro. the sparse design keeps it fast, but all 80b weights still need memory: qwen's 4 bit gguf is about 48 gb, so it wants a well specced mac.

the verdict, by job

• large refactors and migrations: claude opus 5.5 in claude code. highest terminal-bench 4.0 score any vendor has published, and anthropic cites a tester finishing a 680,000 line code migration in less than a day.
• quick edits and everyday pairing: sonnet 5 or gpt-6 sol, both at $2 and $10. pick whichever cli you already like.
• cheap background agents: gemini 3.8 flash is the best value here, at 73.7% on deepswe v1.1 for under a dollar per million input tokens. gpt-6 luna is the cheapest coding model here, at $0.10 and $0.50, if the task is simple. openai reports 66.6% on deepswe v1.1 for it.
• running locally: qwen3 coder next. accept a lower ceiling in exchange for zero api bill.
• self hosting at team scale: glm-5.3, if you have the gpus.
• skip for now: fable 5.1 for coding. in anthropic's table opus 5.5 beats it on terminal-bench 4.0 at 40% of the output price.

why the model is rarely the bottleneck

here is the part the leaderboards leave out. a coding agent fails less often because it cannot write code, and more often because it does not know what you were doing.

it cannot see the stack overflow thread you read yesterday, the slack message where the api shape was agreed, or the error you already chased for two hours. we covered this in giving an ai coding agent memory of your actual work and in what is context engineering. a bigger context window does not help if nothing useful goes in it.

and the models above change every quarter. so your context should not be locked to any one of them.

remynd is a mac app that captures the focused window, runs ocr locally with apple vision, and keeps the searchable index on your mac. you can exclude apps and sites, and recordings are kept for 30 days by default.

it ships an agent window that runs your own claude code or codex cli against that history, with read only access. switch from opus 5.5 to gpt-6 sol next month and the context comes with you. settings also take a custom api endpoint compatible with the openai responses api.

to be precise about privacy: the index lives on your mac, but that does not mean nothing ever leaves it. details are on our security page and in private ai on your mac.

download remynd for mac and let whichever model wins next month start with what you already know.