Claude Opus 5.5 arrives cheaper, with Fable-level claims

Claude Opus 5.5 is Anthropic's attempt to sell frontier-shaped work without frontier-shaped sticker shock. On Sept. 22, 2026 the company posted the model as the first in a new 5.5 family, and TechCrunch's same-day write-up treated the price cut as the news regular buyers would feel first. This is not investment advice, and it is not a leaderboard you should tattoo on a laptop. It is a paraphrase of Anthropic's launch post and one independent rewrite, kept apart where they emphasize different things.

TL;DR - Anthropic says the model performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads at default settings. - Listed token prices are $4 per million input tokens and $20 per million output tokens, versus $5 and $25 for Opus 5. Cache reads are $0.20, which Anthropic calls 60% less than Opus 5. - TechCrunch notes the launch is the first since CEO Dario Amodei argued for pacing capabilities, and that biology and cybersecurity safeguards resemble those on Fable.

What actually happened

Anthropic's post, dated Sept. 22, 2026, introduces the model as a step up from Opus 5 and a cheaper way to get close to Fable-class results. TechCrunch, timestamped that morning, says output tokens are $20 per million, against $25 for the previous Opus, and that other price metrics dropped in a similar way. It also says the model is faster to run because it needs less compute to serve. Anthropic is more numeric: output is more than 30% faster than Opus 5, and a fast mode in Claude Code and the Claude Platform costs $8 per million input tokens and $40 per million output tokens for up to 2.5 times speed.

The cache line matters if you actually run agents. Anthropic's pricing table puts cache reads at $0.20 per million tokens for the new model and $0.50 for Opus 5, cache writes at $5 versus $6.25, and base input at $4 versus $5. The company says cache reads are most of the bill on agentic and coding work, so the 60% cut there is the part it wants you to notice, on top of the "40% less on typical workloads" claim. Those are different measurements. One is a token menu. One is an internal test of typical jobs. Do not collapse them into a single "40% off everything" slogan.

TechCrunch places the release in a short calendar. It says Opus 5 itself shipped on July 24, about two months earlier. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks with many of the same performance, efficiency, and safety changes. TechCrunch repeats that "coming weeks" line. Neither page, in what was opened here, gives those two a date.

Availability, on Anthropic's account, is broad: the model is on the company's own platform as claude-opus-5-5, and also on Amazon Web Services, Google Cloud, and Microsoft Azure. The post also says five-hour usage limits rise on Pro, Max, Team, and seat-based Enterprise plans, and that subscription users get a rate-limit reset they can save and spend when they choose. That is a product knob, not a benchmark.

What the lab wants the scores to prove

Anthropic publishes a table and then warns you not to worship it. On Terminal-Bench 4.0 it lists 66.4% for the new Opus, 55.8% for Fable 5.1, 52.3% for Opus 5, 57.9% for GPT-6 Astra, and 37.3% for GPT-5.6 Sol. On FrontierCode v1.1 Main the figures are 54.4%, 50.3%, 48.0%, 53.3%, and 47.5%. GDPval-AA v2.1 Elos are 1846, 1735, 1708, 1542, and 1588. Footnotes say production safeguards were on. When they intervened, some cyber tasks were finished by Claude Opus 4.8 and some biology tasks by Opus 5, which Anthropic says likely lowers the score. Zapier's AutomationBench result, 40.0%, counts those interventions as failures.

The post also says the real gap versus Fable 5.1 is narrower than the table. That disclaimer belongs next to the chart. Anecdotes are labeled as tests: a 680,000-line migration in under a day, 39 of 40 page-load fixes, a HAProxy rewrite in 9.5 hours versus 12 for Fable at 51% less cost. They are Anthropic's stories, not an independent audit.

Communication is the other half of the pitch. Anthropic says the model puts the important part first, uses less jargon, and follows writing rules more closely. TechCrunch reports the same shift in plainer words: less jargon, important information earlier. A short tester line in the launch post, "it writes the way I do," is marketing color. Treat the price table as the part you can check without trusting a vibe.

Safeguards, and the pacing line

Anthropic calls this its first release since the pacing argument. Named outside evaluators include Frontier Design and METR. On its automated behavioral audit it claims the strongest result it has tested, including fewer hard-to-reverse actions than recent models and better prompt-injection resistance than Opus 5. The same post says the model often suspects it is being evaluated, so the audit is a worse window onto real use. Keep that limit next to the brag.

The lab says the model is comparable to Claude Mythos 5.1 in biology and cybersecurity, so safeguards look like Fable 5.1. Most cyber tasks are re-routed to Opus 4.8. Blocked biology work can go through a Life Sciences Verification Program. Preserved thinking, for distillation defense, applies to API accounts created on or after Aug. 31, 2026. Thinking cannot be switched off. Watermarking is described for the EU AI Act, and zero data retention remains available.

TechCrunch quotes Amodei from a post earlier that month: "I have become convinced that fully addressing the risks requires even more prudence," and then the part about pacing capabilities so risk prevention can keep up. The launch post says public policy has to play a larger role as models get more capable, and that the lab does not assume current measures will be enough for systems that can fully automate AI research. That is a political argument sitting next to a price cut. You can believe the invoice and still notice the argument.

If you are comparing labs rather than invoices, Claude Opus 5.5 is one row, not a single winner. Google's Gemini 4 Argon and a retrieval model like Cohere Embed 5 are different jobs. A coding-agent price cut does not tell you who leads an embedding bench, and an embedding launch does not tell you who should refactor a repo.

What to do as a reader (not a trade)

Check the menu before you repeat the slogan. Claude Opus 5.5 is a price list plus a policy, not a vibe. Claude Opus 5.5 at the default token prices Anthropic printed is $4 in and $20 out, with cache reads at $0.20. The "40% less to run" line is Anthropic's workload test, not a coupon that applies to fast mode. Fast mode is a different row: $8 and $40.

If you already have prompts pinned to Opus 5, read the breaking changes instead of the Elo table. Anthropic says thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to the model and the conversation, and on the Claude API and Google Cloud the older computer-use tool is not accepted. Text between tool calls can come back inside thinking blocks that look empty at the default display setting, so a UI that streamed that text as progress will go quiet until you change the setting. That is the kind of detail that breaks a Tuesday deploy. The leaderboard will not.

Separate safeguards from quality. A refusal, or a silent handoff to an older model, is not the same thing as a wrong answer. If your eval counts a blocked biology or cyber task as a failure, you are measuring the policy as much as the model. Anthropic already says some of its own published scores were reduced by those handoffs. Quote the score with the footnote, or do not quote the score.

Nothing here is a reason to buy or sell a token, a fund, or a seat. The useful reader move is narrower: match the model ID, match the price row you will actually be billed, and match the safeguard class before you let an unattended agent run for hours. Claude Opus 5.5 is a cheaper Opus with a Fable-shaped warning label. Treat it that way.