On 22 September 2026, something unusual happened: Anthropic and OpenAI launched new AI models within roughly an hour of each other. Anthropic released Claude Opus 5.5, the first model in its new Claude 5.5 family. OpenAI countered with GPT-6 Sol and GPT-6 Luna, two cheaper models sitting below its flagship GPT-6 Astra. Both vendors did the same thing — more performance, much lower price. This week, Grok 4.7 and Xiaomi's MiMo-V2.6 also launched, but the Anthropic–OpenAI duel is the one that changes what developers pay for AI.
Here is what the new models actually are, what the benchmark numbers say, and what it means if you write code for a living.
Claude Opus 5.5 is Anthropic's new flagship, replacing Opus 5 after a two-month cadence. The headline is not raw intelligence — it is efficiency:
Safety-wise, it was tested before release by external evaluators METR and Frontier Design, and Anthropic says it attempts to bypass its boundaries about 85% less often than Opus 5. One notable change: "thinking" can no longer be switched off, and it ships with watermarking measures for EU AI Act compliance.
Anthropic published a batch of vendor-run scores. These come from the company itself and secondary coverage, not independent third-party testing — treat them as directional, especially since Anthropic itself warned that at this capability level, "benchmark margins have become a less reliable guide to real-world differences":
| Benchmark | Claude Opus 5.5 | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | vs 57.9% for OpenAI's GPT-6 Astra (as reported by OpenAI) |
| FrontierCode v1.1 | 54.4% | Anthropic's headline coding metric |
| CursorBench 4.0 | 57.8% | vs 41.7% for GPT-5.6 Sol |
| GDPval-AA v2.1 (professional work, 44 occupations) | 1846 Elo | — |
| SWE-bench Pro | 89.9% | vs Claude Fable 5.1 at 81.2% (BenchLM.ai leaderboard, 24 Sep) |
The cost-per-task framing is where it gets interesting for developers: Anthropic claims Opus 5.5 beats GPT-5.6 Sol's top CursorBench score by 11 points at roughly a third of the cost per task, matches GPT-6 Astra on Terminal-Bench 4.0 at about 40% of the cost, and beats Astra's top FrontierCode score at roughly a fifth of the cost per task.
The most concrete claim is a pair of internal engineering tests: translating the HAProxy load balancer from C into Rust, Opus 5.5 finished in 9.5 hours versus 12 for Fable 5.1 at 51% lower cost, with both rewrites passing nearly all regression tests. In a report-writing test where any invented figure or quote failed the quality bar, 16 of 18 Opus 5.5 reports passed — while Fable 5.1 and Opus 5 cleared it zero times. Early testers at GitHub and Deloitte reported fewer tokens per task and higher bug-catch rates in code reviews.
OpenAI's two September 22 releases target the value tier, each roughly half the price of its GPT-5.6 predecessor:
Both sit below the flagship GPT-6 Astra. If you remember GPT-5.6 Sol, I covered why that model is ideal for quick coding answers — here is the free way I use it — and GPT-6 Sol is its direct successor at half the price.
Opus 5.5 looks like the pick for agentic coding work, large migrations, and deep debugging — the tasks where Anthropic's efficiency gains compound. The subscription limit increases also make it attractive for heavy Claude Pro/Max users.
GPT-6 Sol is the better fit for day-to-day coding assistance: fast answers, bug fixes, multi-step reasoning at a fraction of flagship cost. GPT-6 Luna is for background automation and high-volume classification/summarisation where you want the cheapest competent call.
Honest limitations worth knowing:
Two months ago, frontier coding help cost flagship money. Now Opus 5.5 matches the top tier at 40% lower running cost, GPT-6 Sol does serious reasoning at $2/$10, and Luna handles bulk work at pocket change. The practical move for developers: put cheap models on routine work (reviews, summaries, boilerplate), and reserve the expensive calls for the hard problems — exactly the tiered setup that now costs a fraction of what it did in August.
Want to compare the current generation hands-on? Toolxz AI Chat gives you free, no-signup access to frontier models including Claude Fable 5.1 and the GPT-5.6 family — the models whose performance class these new releases are measured against — so you can benchmark the previous generation yourself while the new models roll out everywhere.
When can I use Claude Opus 5.5?
It is live now on the Claude platform, the API, AWS, Google Cloud, and Azure as claude-opus-5-5.
Is GPT-6 Sol the same as GPT-5.6 Sol?
No. GPT-6 Sol is the new September 2026 release, roughly half the price of the GPT-5.6 generation, targeting coding and multi-step reasoning. GPT-5.6 Sol is its predecessor.
Which benchmark should developers trust?
None of them alone. SWE-bench Pro differentiates well on coding (Opus 5.5 leads at 89.9% per the 24 September leaderboard), but OpenAI itself has flagged issues with the benchmark's public split. Use leaderboards to shortlist, then test on your own repos.
Will prices keep falling?
The trend says yes — both vendors cut prices 20–60% this round. But treat today's prices as today's facts and re-check when you build.
Where are these numbers sourced from?
Anthropic's launch announcement and system card (via Unite.AI and Reuters), the AI Stock Wire price-war breakdown, ETV Bharat's launch report, and the BenchLM.ai SWE-bench Pro leaderboard.
Founder of Toolxz (toolxz.com) — 45+ free browser-based tools. I write about practical AI tooling and developer workflows.