Anthropic ships Opus 5.5: Fable-5.1-level on most work, 40% less to run, first Opus with safeguards similar to Fable 5.1's — TechCrunch records a 90-minute gap to OpenAI's own release

3 sources

The launch (Tuesday, September 22, 2026 — calendar-verified). Official positioning: "We're introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." Cadence context: "Claude Opus 5.5 is our first release since we called for pacing the frontier." External evaluators include Frontier Design and METR (official). Media side: per TechCrunch, the model set "a new state-of-the-art in coding and knowledge work performance, according to the company" — the superlative is company self-assessment; and, per Anthropic as relayed by TechCrunch, it "outpaces the larger Fable model in many benchmarks". Safety and safeguards: per Anthropic it is "the strongest-performing model we've tested to date" on the automated behavioral audit (nearly 2,000 scenarios) and the strongest on most measures of honesty; boundary-circumvention attempts came "around 85% less often than Opus 5 or Claude Mythos 5.1", with every attempt low severity and self-reported. Because it is "comparable to Claude Mythos 5.1 in biology and cybersecurity", Anthropic deployed "safeguards similar to those on Claude Fable 5.1" — the first Opus to launch with that similar class of safeguards, which "fall back to another model transparently": most cybersecurity tasks re-route to Opus 4.8, biology work goes through the Life Sciences Verification Program (vetted organizations could apply from September 22), and the Cyber Verification Program expands in the coming weeks. Pricing (three figures, three scopes): $4/$20 per million input/output tokens (20% less than Opus 5); cache reads $0.20 per million (60% less); the 40% figure is total cost on typical workloads at default settings per Anthropic's own "Our tests show" framing; output more than 30% faster; fast mode up to 2.5x speed at $8/$40 per million. Subscriptions: higher five-hour usage limits (Pro, Max, Team, seat-based Enterprise) plus a bankable rate-limit reset. Availability: "now available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure" (official); zero data retention, EU AI Act watermarking, no longer available with "thinking" mode switched off, and preserved thinking for API accounts created on or after August 31, 2026. Efficiency examples (all official-page framing): a 680,000-line code migration in less than a day; a 200,000-line codebase audited and fixed in under three hours where Opus 5 took over 20 hours and 2.5x as many tokens; HAProxy C-to-Rust in 9.5 hours vs 12 for Fable 5.1 at 51% less cost; Deloitte's own quote of 72% of known bugs caught at the lowest effort setting vs Opus 5's 56% at high effort; per Anthropic's page, Walleye Capital reported it noticed and corrected an error in their evaluation instructions, and the page adds, "No other model had caught this error before." The race timeline: per TechCrunch, "Notably, Anthropic released a new version of Opus 5.5 just 90 minutes before OpenAI's release" — note OpenAI's own announcement was unreachable this run (403 on two fetch channels), so that sentence rides the per-TechCrunch chain entirely. Naming note: Anthropic's own hedges ("benchmark margins have become a less reliable guide to real-world differences"; the gap to Fable 5.1 "is narrower than these scores suggest") plus the company-self-assessment attribution on the superlative are the two layers flat versions of this story drop first.

Why now

The pricing took effect Tuesday, September 22 and the model is live on all platforms — for every creator who calls the API or runs Claude Code, this is the rare model story you can see on this month's invoice, which makes it more actionable than any conceptual launch. Second layer, the cadence paradox: this is Anthropic's first release since it "called for pacing the frontier" (the official page's own words) — the tension between calling for a slowdown and shipping a flagship the same quarter is angle gold for commentary content this week. Third layer, per the official page, "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks" — every future 5.5 milestone re-ignites this line, and the creator who explains Opus 5.5's scopes correctly now owns the follow-ups. Fourth layer, the race backdrop (per TechCrunch, 90 minutes between Anthropic's release and OpenAI's) gives the story built-in drama in Tuesday's feeds — but who states the 90 minutes, and whether OpenAI's announcement was opened, is exactly what flat versions erase and comment sections will ask.

Why it is worth publishing

The broadest-audience, hardest-information card of the day: prices (three figures, three scopes), safeguard shape (Fable-class similar safeguards + the Opus 4.8 re-route + two verification-program gates), and availability (all platforms) are all on-screen decision facts, with ready cut-ins for AI-tool reviewers, developer creators, and AI-safety explainers alike. The differentiation play is scope discipline — label the superlative as company self-assessment, keep 40% as the total-cost figure, say "similar class" not "identical", and attach "per TechCrunch" to the 90 minutes — because the flat version erases all four scopes, and each erased scope is one more comment-section own-goal.

Evidence basis

Three sources, all opened with raw HTML fetched this run: Anthropic's official announcement page (its own date label September 22, 2026); TechCrunch's Opus 5.5 report (Russell Brandom, byline 9:30 AM PDT · September 22, 2026; metadata article:published_time 2026-09-22T16:30:07+00:00); TechCrunch's GPT-6 Sol/Luna report (metadata 2026-09-22T18:00:00+00:00, used only for the race sentence). Weekday and date pairs are calendar-verified: September 22 = Tuesday, today = Wednesday, September 23; July 24 carries day granularity per TechCrunch; the Amodei post stays at month granularity ("earlier this month", per TechCrunch) with no day math; "in the coming weeks" (Sonnet 5.5 / Haiku 5.5 / the Cyber Verification expansion) is the official hedge and never becomes a date. Unopened source: OpenAI's own announcement — the announcement URL and two variants returned HTTP 403 on both curl (browser UA) and WebFetch, and openai.com/news returned a JS interstitial; the failures are recorded in the source notes, no source entry was created, no fact on this card depends on the announcement's own text, and all OpenAI-release statements ride the per-TechCrunch chain. Every number carries its holder and scope: 40% (total cost, typical workloads, default settings, Anthropic's own test framing) / 20% (per-token drop) / 60% (cache reads) / more-than-30% (output speed) / 85% (reduction in boundary-circumvention attempts) / 72% vs 56% (Deloitte's own quote) / 51% (HAProxy cost) / 2.5x (fast-mode speed and the token comparison — two different referents, each labeled) / 90 minutes (per TechCrunch); the official page states no percentage for the five-hour usage-limit increase, and none is imported.

“By Anthropic's own account, Opus 5.5 performs at Fable 5.1's level on most work while costing 40% less to run than Opus 5 — the lab's first release since it called for pacing the frontier.”

Angle

Frame it as a 'three figures, three scopes' model-release explainer in four beats. Beat one, how to read the price: $4/$20 per million input/output tokens (20% less than Opus 5), cache reads $0.20 (60% less), and the 40% figure is total cost on typical workloads at default settings — the three numbers must not be blended into one claim. Beat two, how to say the capability: per TechCrunch's own sentence the state-of-the-art call is company self-assessment, and Anthropic itself notes benchmark margins are 'a less reliable guide' and the real gap to Fable 5.1 is narrower than the scores — every conclusion carries its holder. Beat three, what the safeguards actually are: the first Opus shipping with Fable-5.1-class safeguards (cybersecurity, biology, distillation), with most cybersecurity tasks re-routed to Opus 4.8, biology behind the Life Sciences Verification Program, and the Cyber Verification Program expanding in the coming weeks — say 'similar class', not 'identical'; say 'gated', not 'banned'. Beat four, cadence and the race: this is Anthropic's first release since calling for pacing the frontier; per TechCrunch, OpenAI released 90 minutes later — keep the attribution on the race.

Format

Short talking-head video

Demo idea

A 'three figures, three scopes' price card: left column three numbers ($4/$20, $0.20, 40%), right column each scope (per-million-token prices, cache reads, total cost on typical workloads), each row badged 'Anthropic's official pricing table'; closing card: 'the 40% is a total-cost figure, not the per-token drop.' Second card, a timeline strip: July 24, Opus 5 (TechCrunch day-granularity badge) → earlier this month, Amodei's post (month-granularity badge) → September 22, Opus 5.5 (official + TechCrunch dual-source badge) → same day, 90 minutes later, OpenAI's release (single-source 'per TechCrunch' badge) → coming weeks, Sonnet 5.5 / Haiku 5.5 (official-hedge badge, no date).

Platform notes

The superlative appears only as 'company self-assessment' or 'company self-assessment, as relayed by TechCrunch' — never as an independent benchmark verdict; read Anthropic's own two hedges (benchmark margins are a less reliable guide; the real gap to Fable 5.1 is narrower) alongside it. Loaded-label discipline: the official page's 'attempting to escape a sandbox' describes behaviors from past incidents that Opus 5.5 improves on — never write that Opus 5.5 escaped, hacked, or went rogue; its own stat is the reverse (boundary-circumvention attempts down around 85%, every attempt low severity and self-reported). Safeguards are 'similar class to Fable 5.1' — never 'identical'; name the Opus 4.8 re-route and the two verification-program gates — neither 'open to everyone' nor 'banned' is accurate. Keep the three price scopes separate: $4/$20 and $0.20 are per million tokens, 40% is the total-cost figure on typical workloads at default settings (Anthropic's own test framing), and fast mode is a separate, pricier tier ($8/$40) — never drop 'per million tokens'. The race sentence always carries 'per TechCrunch' — OpenAI's announcement was not opened this run, and OpenAI's own model claims (price halving, factuality) stay out of this card. Timing: the announcement date is Tuesday, September 22, 2026 and this card publishes September 23 — announcement-relative 'today' words only inside attributed quotes; 'coming weeks' never becomes a date; 'earlier this month' takes no day math; 'just two months after the release of Opus 5 on July 24' is TechCrunch's own span and stays attributed. Deloitte's 72%-vs-56% is Deloitte's own quote, and the Walleye error-catch is Anthropic's page prose about Walleye's report — name the holder of each when citing.

Usable claims

Anthropic released Claude Opus 5.5 on Tuesday, September 22, 2026 (calendar-verified). Per Anthropic: "We're introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." Per Anthropic: "Claude Opus 5.5 is our first release since we called for pacing the frontier." And: "It was tested before release by external evaluators, including Frontier Design and METR." Per TechCrunch, the new model was "released on Tuesday, setting a new state-of-the-art in coding and knowledge work performance, according to the company", and "Notably, Anthropic says, the release outpaces the larger Fable model in many benchmarks and succeeded in a number of informal tasks that Fable failed to complete." On safety, per Anthropic: "On our automated behavioral audit, the most comprehensive alignment test we run, Opus 5.5 is the strongest-performing model we've tested to date." The audit assesses Claude across nearly 2,000 scenarios, and per Anthropic Opus 5.5 is "also our strongest model on most measures of honesty." Per Anthropic: "In a new evaluation designed to test a model's propensity to cross containment boundaries, Opus 5.5 attempted to circumvent boundaries around 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt it made was low severity and self-reported." On safeguards, per Anthropic: "Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we're deploying it with safeguards similar to those on Claude Fable 5.1." And: "Opus 5.5 is the first Opus model to launch with a similar class of safeguards to Fable 5.1 on cybersecurity, biology, and distillation, all of which fall back to another model transparently." Per Anthropic on cyber: "Users will be able to identify and fix bugs in their code as part of the routine software development lifecycle, but most cybersecurity tasks will be re-routed to Opus 4.8." Per Anthropic: "Vetted organizations can apply today to our Life Sciences Verification Program to use Opus 5.5 for biology research." And: "In the coming weeks we will also be expanding access to our Cyber Verification Program, and verified cybersecurity practitioners will be able to use Opus 5.5 for their work." Per TechCrunch: "Those safeguards limit how much the models can be used to discover exploits in compiled programs or developing recognizable biological weapons, among other tasks." On availability, per Anthropic: "Claude Opus 5.5 is now available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure." And: "Like previous Opus models, Opus 5.5 is available with zero data retention." It ships with the same watermarking measures as Fable 5.1 to comply with the EU AI Act, and per Anthropic: "It is also no longer available with "thinking" mode switched off, as we describe here." Per Anthropic: "Opus 5.5 is launching with preserved thinking, the anti-distillation safeguard we introduced with Fable 5.1." That safeguard "stops API users from editing Claude's prior context in an attempt to extract Claude's reasoning" and "applies to Fable 5.1 and Opus 5.5 for API accounts created on or after August 31, 2026."
On pricing, per Anthropic: "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that. Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads." Per Anthropic: "Input and output tokens are $4 and $20 per million, 20% less than Opus 5." Per Anthropic: "Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5." And per Anthropic: "Opus 5.5 also generates output more than 30% faster than Opus 5." Per Anthropic: "Fast mode for Opus 5.5 is also available in Claude Code and the Claude Platform with up to 2.5x speed. It costs $8 per million input tokens and $40 per million output tokens." Per TechCrunch: "Output tokens will be charged at $20 per million tokens for Opus 5.5, compared to $25 for the previous model." And: "Other metrics have similar price drops." Per Anthropic: "In addition to the price drop, we're increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. We're also providing subscription users a rate limit reset, which you can now save and use whenever you choose." Per Anthropic: "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same improvements to performance, efficiency, and safety." Per TechCrunch: "The launch comes just two months after the release of Opus 5 on July 24." — both endpoints of that span are day-precise, and the "just two months" phrasing is TechCrunch's own. On communication, per Anthropic: "It puts the most important information up front, and its style makes it a better work partner over long sessions." And, as one early tester put it per Anthropic: "it writes the way I do." Per TechCrunch: "The new version also makes significant changes to how Opus communicates, with the Opus 5.5 less likely to use jargon and more likely to put important information at the start of its messages." On efficiency, per Anthropic: "One tester completed a 680,000-line code migration in less than a day—work that would have taken an engineering team weeks." And: "An early tester used it to audit and fix a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5x as many tokens." On Anthropic's HAProxy test: "Both rewrites passed nearly all of HAProxy's own regression tests, but Opus 5.5 finished in 9.5 hours compared to 12 for Fable 5.1, and cost 51% less." Per Anthropic: "At its default effort level on FrontierCode, it beats GPT-6 Astra at roughly 20% of the cost per task." Deloitte Consulting LLP, an early tester, states in its own quote on Anthropic's page: "Even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5's 56% at high effort, with fewer false alarms and a fraction of the output." And per Anthropic's page, Walleye Capital reported that Opus 5.5 noticed an error in their evaluation instructions and corrected for it; the page adds, "No other model had caught this error before." On the competitive timing, per TechCrunch: "Notably, Anthropic released a new version of Opus 5.5 just 90 minutes before OpenAI's release, reflecting the intense competition between the two companies." OpenAI's own announcement was unreachable this run (HTTP 403 across two fetch channels), so the race sentence rides TechCrunch's opened report alone.

Evidence pipeline

From the news

Breakdown

One model launch packed with three kinds of numbers whose scopes get flattened first — read it in layers. Layer one, the launch itself: the official positioning is 'performs at the level of Claude Fable 5.1 on most work' at '40% less to run than Opus 5', and 'our first release since we called for pacing the frontier'; TechCrunch's state-of-the-art sentence carries its own 'according to the company' attribution. Layer two, number scopes: the 40% is total cost on typical workloads at default settings (the official page's own 'Our tests show' framing), not the per-token drop — the per-token scopes are $4/$20 per million (20% less) and $0.20 cache reads (60% less); the 85% is the drop in boundary-circumvention ATTEMPTS, every attempt low severity and self-reported — not any kind of failure rate; the 90 minutes is a single-source TechCrunch claim — OpenAI's own announcement was unreachable this run (403 on both fetch channels), so the sentence must carry 'per TechCrunch'. Layer three, safeguard shape: the official wording is 'safeguards similar to those on Claude Fable 5.1' / 'a similar class of safeguards', and they 'fall back to another model transparently' — most cybersecurity tasks re-route to Opus 4.8; the access gates are real (Life Sciences Verification Program open to vetted organizations since September 22; the Cyber Verification Program expanding in the coming weeks), so both 'fully open' and 'banned' are wrong. Layer four, the publisher's own hedges: benchmark margins are 'a less reliable guide to real-world differences' and the real gap to Fable 5.1 'is narrower than these scores suggest' — the company cooling its own benchmarks is the most quotable material in the story. Layer five, tester-number attribution: Deloitte's 72%-vs-56% is Deloitte's own quote, and the Walleye error-catch is Anthropic's page prose about Walleye's report — both hosted on Anthropic's own page; name the holder, never platform them as independent conclusions. Editor's rules: superlatives carry holders, three prices carry three scopes, loaded labels attach only to the past-incident behaviors, the race carries its single-source tag, and the official hedges get read aloud.

Sources

Risks

Before publishing, re-check each layer: every benchmark or superlative with its holder (company self-assessment vs tester quote vs outlet framing); every loaded label attached to the past-incident behaviors it describes, never to Opus 5.5; safeguard wording as 'similar class', with the Opus 4.8 re-route and the two verification-program gates named; each price figure with its scope (per-token vs cache reads vs total cost) and 'per million tokens' intact; the 90-minute race only inside the 'per TechCrunch' chain with no OpenAI model claims imported; and every date either calendar-verified (September 22 = Tuesday) or left at the source's own granularity. If your script compresses any of these, cut the detail rather than round it.

Demo ideas

'Three figures, three scopes' price card ($4/$20, $0.20, and 40% side by side with their scopes, each row carrying the official-pricing-table badge), closing on 'the 40% is a total-cost figure, not the per-token drop'
Timeline strip: July 24, Opus 5 (TechCrunch day-granularity badge) → earlier this month, Amodei's post (month-granularity badge) → September 22, Opus 5.5 launch (official + TechCrunch dual-source badge) → same day, 90 minutes later, OpenAI's release ('per TechCrunch' single-source badge) → coming weeks, Sonnet 5.5 / Haiku 5.5 (official-hedge badge, no date)