Last updated: September 2026. An early tester used Claude Opus 5.5 to migrate a 680,000-line codebase in less than a day — work that would keep an engineering team busy for weeks. That is the headline behind Anthropic's September 22, 2026 launch. Claude Opus 5.5 is the first model in the Claude 5.5 family: it works at Claude Fable 5.1 level on most tasks, costs 40% less than Opus 5 on typical workloads, and writes more clearly than the model it replaces.
The announcement landed about 90 minutes before OpenAI shipped its own GPT-6 Sol and Luna models. Two frontier releases in the same afternoon is unusual even by 2026 standards, and it makes direct comparison unavoidable — a comparison we handle honestly below, including where Opus 5.5 loses.
Source: Anthropic
What Is Actually New Here
Opus 5.5 is not just a price cut with a new badge. Anthropic describes a performance jump over Opus 5 across coding, agentic work, and communication, with three changes that matter for daily use.
First, raw capability. The early-access results include the 680,000-line migration, plus a striking consistency test: Opus 5.5 improved a web app's page load times in 39 of 40 attempts, where Opus 5 produced smaller improvements that also changed app behavior. One tester built a complete game project from a single prompt, and Anthropic says it beat every other Claude model on graphics quality and finishing touches.
Second, writing quality. Users criticized Opus 5 for dense, hard-to-follow prose. Opus 5.5 is tuned to write more clearly and to lead with the most important information. One tester's verdict: "it writes the way I do."
Third, the economics. Anthropic rebuilt the cost structure around how people actually use coding and agent tools, where cached reads dominate the bill. The result is the 40% workload-level saving, broken down in the pricing table below.
The Official Benchmark Table
Here are Anthropic's published numbers, exactly as reported. Note where GPT-6 Astra wins — we discuss that in full in the limitations section.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 (knowledge) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench (business automation) | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity's Last Exam (with tools) | 67.7% | 65.6% | 63.6% | 57.2% | — |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | low |
The table shows Opus 5.5 ahead of every other Claude model listed, reaching Fable 5.1's frontier level at a much lower price. For context on the previous generation, our Claude Opus 5 launch coverage has the full story.
Pricing and Speed: The Numbers That Matter
The pricing shift is the most aggressive part of this release, especially for anyone running agents or coding assistants around the clock.
| Pricing item | Opus 5.5 | vs Opus 5 |
|---|---|---|
| Input tokens | $4 / million | 20% cheaper |
| Output tokens | $20 / million | 20% cheaper |
| Cache reads | $0.20 / million | 60% cheaper |
| Typical workloads | — | 40% cheaper overall |
| Output speed | — | 30%+ faster |
The cache-read line is the sleeper story. Anthropic notes that cached reads make up the majority of costs for agent and coding workloads, because agents repeatedly re-read the same context. Cutting that line by 60% is what turns the headline token prices into a real 40% reduction on the bill. If you are comparing this against OpenAI's half-price GPT-6 generation, our GPT-6 Sol and Luna launch report has the other side of that calculation.
Source: Anthropic Newsroom
Safety and External Testing
Opus 5.5 is the first release under Anthropic's "pacing the frontier" initiative, and the safety work here is unusually detailed. Before launch, the model was assessed by external evaluators including Frontier Design and METR, not just Anthropic's internal teams.
On Anthropic's most comprehensive internal alignment test, an automated behavioral audit, Opus 5.5 scored better than any previous model. Anthropic reports it is less inclined than recent models to take hard-to-reverse actions or to exceed the permissions it has been granted, and more resistant than Opus 5 to prompt injection attacks. The full technical detail is in the Opus 5.5 System Card.
The capability picture required extra safeguards too. Anthropic rates Opus 5.5 as comparable to Claude Mythos 5.1 in biology and cybersecurity, so it ships with the same protections as Fable 5.1, plus a Life Sciences Verification program for accredited research organizations. An expanded Cyber Verification program for accredited security professionals follows in the coming weeks.
What It Means for Pro, Max, Team, and Enterprise Users
The subscription news is just as practical as the API pricing. Anthropic raised the usage limits for every 5-hour window on the Pro, Max, Team, and Enterprise plans, calculated per seat. Heavy users who kept hitting the ceiling mid-task get real headroom back.
The smarter change is the bankable rate-limit reset. Instead of a reset that fires automatically whether you need it or not, subscribers can now save it and trigger it when they actually want it — for example, right before a long evening coding session. It is a small feature with a large effect on how the plan feels in daily use.
For teams evaluating where Opus 5.5 fits, Truescho's AI tools directory tracks the current generation of models side by side.
What Comes Next: Sonnet and Haiku Join the Family
Opus 5.5 is only the opening move. Anthropic confirmed that Claude Sonnet 5.5 and Claude Haiku 5.5 are coming in the coming weeks, carrying the same family-wide improvements: the lower pricing structure, the faster output, and the clearer writing style.
That roadmap matters for planning. If you run a tiered setup — Opus for hard architecture work, Sonnet for everyday implementation, Haiku for high-volume classification — there is little reason to lock yourself into a competitor's stack this quarter. Waiting a few weeks gets you the same cost and quality gains across every tier, inside one billing account.
It also signals how the "pacing the frontier" initiative will work in practice: staged releases, external evaluation before launch, and published safety documentation for each model. Whether that cadence holds is an open question, but the first data point is encouraging.
The Honest Limitations
Three things keep this launch from being a clean sweep.
First, GPT-6 Astra beats Opus 5.5 on two published benchmarks. Astra scores 41.4% versus 40.0% on AutomationBench, and 64.6% versus 58.7% on Terminal-Bench-Science. If business automation or scientific terminal work is your exact workload, Astra deserves a head-to-head test before you commit — our GPT-6 Astra launch report covers its strengths in detail.
Second, Anthropic itself warns against over-reading the table. The company says that at these capability levels, benchmark margins have become "a less reliable indicator" of real-world differences, and that the true gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. When the vendor tells you its own wins are smaller than they look, believe it.
Third, these are Anthropic's numbers, produced by Anthropic, on evaluations Anthropic selected. The external testing helps on safety, but performance comparisons are still vendor-supplied until independent labs publish.
FAQ
How much does Claude Opus 5.5 cost per million tokens?
Input costs $4 per million tokens and output costs $20 per million. Cache reads cost $0.20 per million. Anthropic says typical workloads cost 40% less than Opus 5 overall.
Is Claude Opus 5.5 better than GPT-6 Astra for coding?
On agentic coding benchmarks, Opus 5.5 leads: 66.4% on Terminal-Bench 4.0 versus Astra's 57.9%. But Astra wins on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%). The honest answer: test both on your own codebase.
What are the improvements of Opus 5.5 over Opus 5?
Cheaper and faster pricing (20% lower token prices, 60% cheaper cache reads, 40% cheaper typical workloads, 30%+ faster output), a large performance jump on coding and knowledge benchmarks, clearer writing that addresses Opus 5 feedback, and better safety results including stronger prompt injection resistance.
When will Claude Sonnet 5.5 and Haiku 5.5 be released?
Anthropic says both are coming in the coming weeks, with the same family-wide improvements as Opus 5.5.
Did Anthropic increase rate limits for Pro and Max users?
Yes. Pro, Max, Team, and Enterprise plans get higher limits per 5-hour window, and Anthropic added a bankable rate-limit reset that subscribers can trigger whenever they choose.
Our Verdict
Claude Opus 5.5 is the rare release where the marketing and the mechanics agree. Fable-level output at Opus pricing, cache reads cut by 60%, and a 680,000-line migration done before dinner make a strong case that this is the best value Anthropic has ever shipped — with genuinely improved safety documentation behind it. It does not beat GPT-6 Astra everywhere, and Anthropic admits its own benchmark margins are thinner than they look. But for coding, agents, and long-form work on a budget, Opus 5.5 is now the model to beat. While you rethink your AI workflow, browse Truescho's open opportunities to see what new skills this model generation makes worth learning.
Sources
- Anthropic — Introducing Claude Opus 5.5 — the official announcement with benchmarks, pricing, and safety details
- Anthropic Newsroom — official statements on the Claude 5.5 family rollout
- OpenAI — Introducing GPT-6 Sol and Luna — the competing launch announced 90 minutes later