Claude Opus 5.5 Is Out: Price, Specs, Every Leak Graded

§ 01What Anthropic released#

The models overview now opens with Opus 5.5 as the model to start with for most workloads, and moves Claude Opus 5 to the legacy list beside Fable 5 and Opus 4.8. Fable 5.1 keeps the recommendation for the most demanding reasoning and long-horizon agentic work, at $10 and $50.

§ 02The launch table, as Anthropic printed it#

Anthropic’s own caveat sits under the table: “at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences”, and in its own use the gap to Fable 5.1 is narrower than the scores suggest. With that said, here is the table.

Opus 5.5 leads 7 of the nine rows; GPT-6 Astra keeps AutomationBench and Terminal-Bench-Science. Three footnotes matter. Terminal-Bench 4.0 is reported at xhigh effort for Opus 5.5 and at high for GPT-6 Astra, each model’s best; the GPT figures are OpenAI’s own. AutomationBench was run by Zapier without fallback models, so safeguard interventions counted as failures. And when Opus 5.5’s safeguards intervened during evaluation, cybersecurity tasks were completed by Opus 4.8 and biology tasks by Opus 5, which Anthropic says likely lowers its scores. On GDPval-AA, Opus 5.5 gains 138 Elo on Opus 5 and 111 on Fable 5.1.

The efficiency claim is the one Anthropic leans on hardest: at its default effort on FrontierCode it beats GPT-6 Astra at roughly 20 percent of the cost per task, on Terminal-Bench 4.0 it matches Astra at about 40 percent of the cost, and “Opus 5.5 also generates output more than 30% faster than Opus 5.” An internal HAProxy rewrite from C to Rust finished in 9.5 hours against 12 for Fable 5.1 at 51 percent lower cost.

§ 03Pricing: the leaked sheet was exact#

At 15:30 UTC on September 20, @synaptized posted four numbers. At 16:45 UTC on September 22, Anthropic’s pricing page carried the same four.

The cache read is the row that changes agentic economics: at $0.20 it is 0.05x the input price where Opus 5’s was 0.1x, and Anthropic notes cache reads make up the majority of agentic and coding costs. Alongside the price, Anthropic raised the five-hour usage limits on Pro, Max and Team plans and gave subscribers a rate-limit reset they can save and use when they choose.

§ 04Every claim from September 20, graded#

Four of the nine held, one was contradicted, three cannot be checked against any document, and one was false before the day was out. The leaker’s identifier and date and the price account’s figures were right; the render that carried the story’s first hours told you nothing about the model.

§ 05What changed for developers#

Anthropic’s own notes list four breaking changes from Opus 5. Thinking cannot be disabled: a request with thinking set to disabled, or a manual thinking budget, returns a 400 invalid_request_error, and the effort parameter is the control. “Claude Opus 5.5 doesn’t support forced tool use.” Setting tool_choice to any or to a named tool returns a 400; auto plus strict tool use is the replacement. Thinking blocks record the model that produced them: Opus 5.5 reads blocks from Opus 5 and earlier Opus, Sonnet and Haiku models but not from Fable or Mythos, and only Fable 5.1 and Mythos 5.1 read its blocks. And the computer_20251124 tool returns an error on the Claude API and Google Cloud; the computer_toolset_20260801 toolset is the path.

Two behavior changes arrive without a code change. The default effort is medium where Opus 5 ran at high, so a request that omits effort thinks less than before, and at the same effort the model thinks more per turn, most at xhigh and max. Text between tool calls now arrives as progress-update thinking blocks rather than text blocks, so an application streaming those notes to users goes quiet between calls unless it changes its display setting.

§ 06Safeguards#

Anthropic frames this as its first release since its CEO’s “We Must Pace the Frontier” essay, tested before release by Frontier Design and METR. Because the model is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to Fable 5.1’s: “most cybersecurity tasks will be re-routed to Opus 4.8”, with a Cyber Verification Program expanding to Opus 5.5 in the coming weeks, and biology work beyond the safeguards through the Life Sciences Verification Program. Preserved thinking, the anti-distillation measure introduced with Fable 5.1, applies to API accounts created on or after August 31, 2026. On Anthropic’s automated behavioral audit of nearly 2,000 scenarios, it reports Opus 5.5 as its strongest model to date on most measures of misaligned behavior, attempting to cross containment boundaries about 85 percent less often than Opus 5, while noting the model often suspects it is being evaluated.

On September 24, 2026, Anthropic changed how refusals are billed, on every platform. A refusal that arrives before any output is now charged like any other request, at the rates of the model that ran it, when its stop_details.category is bio, frontier_llm or reasoning_extraction, “the categories where we measure low volumes of false positives”. Mid-stream refusals were already billed. Refusals before any output in the other categories, cyber among them, are still free, and fallback credit is unchanged. For a team sending biology work to Opus 5.5, a refusal is now a line on the bill rather than a free miss.

§ 07The cadence, closed#

Opus 5.5 landed 21 days after Fable 5.1 and 60 days after Opus 5, the short end of the year’s range and exactly where the September 20 leak placed it. “Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks”, per the announcement, which would make this the first Anthropic generation to ship Opus first.

§ 08What the market said#

Polymarket’s next-Claude-Opus market priced September 24 at 73.5 percent and September 30 at 89 percent when we read it on September 20 at 19:12 UTC. The September 24 question resolves yes. The market moved on the leaker’s posts, and the leaker was right; that is a fact about this leak, not a rule about the next one.

Two adoption signals arrived on October 3, 2026. At 11:00 UTC Theo, who builds the T3 Code app, posted: “Opus 5.5 is the first model we’ve seen break the 50% traffic threshold in T3 Code.” His unit is prompts (“Literally half of all prompts go to Opus”); a reply asked whether tokens and spend look the same, and the post does not say. The same morning Google’s Antigravity listed Claude Opus 5.5 and Claude Sonnet 5.5, both with thinking, on its models page for Google AI Ultra and for non-trial Google AI Pro subscriptions, and not for the Free, Google AI Plus or Enterprise plans. Antigravity’s changelog carries no entry for the addition; the same page says Claude Opus 4.6 and Sonnet 4.6 leave Antigravity on November 2, 2026. One app’s prompt share and one IDE’s plan table are not a market share, but both point the same way two weeks after release.

§ 09Where CellCog stands#

Our conflict, declared: we build CellCog, a platform where a business hires AI employees that keep memory and work as a team, and we choose the models underneath them. On September 22, 2026, the day it shipped, every CellCog tier that ran Claude Fable 5.1 moved to Opus 5.5: Agent, Creative and Team, Core and Max. Flash followed on September 26, moving from Gemini 3.8 Flash to Opus 5.5 at low effort, so every CellCog tier now runs Opus 5.5. The receipt is our day-one product update, with what changed and what did not.

§ 10What we are watching for#

• The system card, published with the release, for the evaluation detail behind the launch table.
• Independent scores: Artificial Analysis, the Terminal-Bench and OSWorld leaderboards, and the DeepResearch Bench leaderboard, where Anthropic’s own table is not the judge.
• Claude Sonnet 5.5 and Claude Haiku 5.5, which now have their own pages: Sonnet 5.5 and Haiku 5.5.
• Whether more coding tools publish a model share the way T3 Code did, and whether Antigravity opens Opus 5.5 to Enterprise.
• Whether the codename claude-wafer-eap ever appears in an Anthropic document, or in an API error.
• How the model behaves across a full working day of AI-employee work on our own tiers, now that it runs them.

§ 11Update log#

This is a living page; when the story moves, the update lands here.

As of October 4, 2026, 07:06 UTC: adoption. Theo’s October 3 post on T3 Code prompt share and Antigravity’s models page (Opus 5.5 and Sonnet 5.5 on Ultra and non-trial Pro) folded into What the market said; the Flash routing line corrected (Opus 5.5 at low effort since September 26).

As of September 25, 2026, 02:35 UTC: Anthropic’s platform release notes (September 24 entry) extend billing to refusals that arrive before any output in the bio, frontier_llm and reasoning_extraction categories; folded into Safeguards.

As of September 22, 2026, 19:25 UTC: CellCog moved every tier that ran Claude Fable 5.1 to Opus 5.5 the same day; the receipt is the day-one product update linked above.

As of September 22, 2026, 16:50 UTC: released. Anthropic’s post at 16:31 UTC and the announcement page, the platform release notes (September 22 entry), the models overview, the pricing page, the migration notes and OpenRouter’s catalog all read within fifteen minutes of the post. The tracker became this record: specs and prices from the documentation, the launch table as printed, every September 20 claim graded. The Opus 5.2 tracker keeps the earlier chain.

As of September 20, 2026, 19:15 UTC: page opened. Every X post read directly in a browser and its time computed from the post id; Anthropic’s model documentation, release notes, pricing page, status page and OpenRouter’s Anthropic catalog read at 19:10 UTC; Polymarket read at 19:12 UTC. No official movement to log.

§ 12Sources#

• Anthropic, Introducing Claude Opus 5.5, September 22, 2026, and the @AnthropicAI post, 16:31 UTC, read directly.
• Anthropic, What’s new in Claude Opus 5.5, models overview, pricing, platform release notes, all read September 22, 2026, 16:45 UTC; OpenRouter models API, same time.
• X posts of September 20, 2026, read directly, times computed from the post ids: @lyraxana, 13:43 UTC; @synaptized, 15:30 UTC and 15:32 UTC; @HarshithLucky3, 12:38 UTC, 13:25 UTC, 13:47 UTC and the 19:00 UTC retraction.
• Anthropic release posts: Claude Opus 5, July 24, 2026; Claude Fable 5.1 and Mythos 5.1, September 1, 2026; Claude Fable 5 and Mythos 5, June 9, 2026; Claude Sonnet 5, June 30, 2026.
• Polymarket, Next Claude Opus released by…?, read September 20, 2026, 19:12 UTC.
• Anthropic, platform release notes, September 24, 2026 entry, and How refusals are billed, read September 25, 2026, 02:31 UTC.
• @theo, October 3, 2026, 11:00 UTC, read directly. Google, Antigravity models and Antigravity changelog, read October 4, 2026, 07:06 UTC.
• CellCog routing statements reflect our configuration on September 22, 2026, and the Flash change of September 26, 2026.