Anthropic Cuts Opus 5.5 Costs to Slow Open-Weight Defections

Anthropic launched Claude Opus 5.5 on September 22, 2026, at $4 and $20 per million tokens, 20% below Opus 5. The company said typical workloads billed by token now run 40% cheaper, and that the model performs at the level of Claude Fable 5.1 on most work.

That 40% figure is not the list-price cut. It is Anthropic’s estimate of a cheaper rate plus shorter answers, arriving as software teams were already sending most production tokens to open-weight models, including Chinese ones, to escape frontier bills.

Claude Opus 5.5 Drops List Prices by 20%

The first 5.5-family model is live as claude-opus-5-5 on the Claude Platform and in Amazon Web Services, Google Cloud, and Microsoft Foundry. Paid Claude plans get it too, and Anthropic made it the default Opus in Claude Code. Input and output on the $4 per million input tokens sheet sit 20% under Opus 5’s $5 and $25, the first Opus list cut after the July 24, 2026 Opus 5 launch held the old number.

Anthropic called it “the new leading model,” and said early testers saw large jumps on their most complex work. One tester finished a 680,000-line code migration in less than a day, work the company said would have taken an engineering team weeks. Output, at standard speed, is more than 30% faster than Opus 5.

THE 5.5 LIST SHEET

Fable 5.1 still costs more than twice Opus 5.5 on fresh input and output. GPT-6 Sol, OpenAI’s cheaper GPT-6 tier, undercuts the new Opus on the same two lines. Astra matches Fable’s $10 and $50 sticker and is the expensive OpenAI comparison Anthropic wants buyers to make.

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 for most tasks, and costs 40% less to run than Opus 5. pic.twitter.com/Q9C2VKQ79f

— Claude (@claudeai) September 22, 2026

What a 40% Typical Workload Cut Means

Anthropic’s 40% claim is a blended estimate at default settings, not a coupon on every request. List rates fell 20%. The rest has to come from using fewer tokens per finished job. The company’s own tests say Opus 5.5 reaches the same result with about 20% fewer input and output tokens, and writes about 40% less verbose output.

A lower rate times a shorter answer compounds. FinOps LLM, which prices LLM traffic, walked the arithmetic as 0.8 × 0.8 = 0.64, or about 36% to 40% off depending on the mix. On an identical agent step of 200,000 input tokens with 90% served from cache and 20,000 output tokens, the bill falls from $0.690 on Opus 5 to $0.516 on Opus 5.5, a 25% drop before any token saving. If the 20% token cut holds, that step lands near $0.41, which is where Anthropic’s 40% lives.

Effort settings can wipe that out. Independent tester Artificial Analysis put Opus 5.5 at a score of 58 on its Intelligence Index at max effort, the highest it has measured, and gave the model the top spot on the Intelligence Index. At that same max setting it used about 119,000 output tokens per Index task, against about 73,000 for Opus 5, so cost per task can end up level if a team always cranks the dial. The 40% story is a default-settings story. It is not a max-effort story.

Cache Reads Are Where Agent Bills Get Paid

The line that moved most is the one most launch posts bury. Cache reads fall from $0.50 to $0.20 per million tokens, a 60% cut. Anthropic said those reads make up the majority of costs in agentic and coding work, because long jobs reread the same repo, prompt, and tool trace. Against uncached input at $4, a hit now costs 5% of a fresh token, a 95% discount, up from 90% on earlier Opus models.

CACHE READ

• The mechanic: The API stores a prompt prefix and later rereads it instead of billing the full input rate again.
• The new hit rate: Opus 5.5 cache hits are 0.05× base input, $0.20 per million, versus 0.1× on most other Claude models.
• The miss penalty: A broken prefix that misses cache now costs 20 times a hit, so prompt-template churn got more expensive, not less.

Five-minute cache writes are $5, down from $6.25. One-hour writes are $8. Batch input and output are $2 and $10, half the standard rate. Fast mode, a research preview on the first-party Claude API only, doubles the sticker to $8 and $40 for up to 2.5× speed, and it is not on AWS, Google Cloud, or Microsoft Foundry. US-only inference is 1.1× on every token category.

WHAT ELSE MOVED WITH THE CUT

• Session limits: Five-hour limits on Pro, Max, Team, and seat-based Enterprise rise 20%, and the lower token price means those limits go 25% further.
• Saved reset: Pro, Max, and Team users get a rate-limit reset in Settings that can be applied any time until October 22, 2026.
• Always-on thinking: Thinking cannot be turned off. Medium is the default, and teams change quality with the effort setting.
• Cloud ID: Developers call claude-opus-5-5 through the Claude API and the three major clouds, with a 1 million token context window.

The switch is not free in engineering time. Forced tool-choice values that used to work now return errors, computer-use calls need a new tool version, and thinking blocks stay bound to the model and the conversation. A cheaper Opus still has a migration bill.

The Token Majority Already Left Frontier APIs

The commercial pressure is already visible in traffic, not in a future slide. On Vercel’s AI Gateway, open-weight models processed 56% of production tokens in August 2026, up from 36% in July and from under 10% in December 2025. Those models were only about 14% of estimated spending, which is the point: volume moved because the unit price moved.

WHERE THE TOKENS WENT

• August share: Open-weight models handled 56% of Vercel gateway tokens, and 62% on one record day cited by chief executive Guillermo Rauch.
• Spend share: The same class was about 14% of estimated spending, so the cheap models took the work, not the revenue.
• Gap in months: A Mozilla report dated September 15, 2026 put the performance gap between US frontier systems and the best Chinese open-weight models at 4.4 months.
• DoorDash split: DoorDash has used Moonshot’s Kimi for routine work and reserved Fable for harder tasks that would take human experts longer.

Coinbase made the swap public on June 29, 2026. Chief executive Brian Armstrong said the exchange defaulted engineers to GLM 5.2 from Zhipu and Kimi 2.7 from Moonshot through an internal gateway, cutting internal AI spend by nearly half even as token use hit records. Together AI chief executive Vipul Ved Prakash has put the best open-weight models at 10 to 60 times cheaper than closed equivalents. Anthropic’s 40% typical cut is a real move inside a closed API. It is not in that 10-to-60-times band.

Vercel also said nine in 10 teams using Fable reduced that usage, and that more of those customers moved to Opus than to any other model. Opus 5.5 is Anthropic trying to catch the people who already decided Fable was twice the price for most production work.

How Opus 5.5 Compares With GPT-6 Sol

On Anthropic’s own table, Opus 5.5 leads several agent and knowledge-work scores. Terminal-Bench 4.0 is 66.4% against 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. FrontierCode v1.1 is 54.4% against 50.3%, 48.0%, 53.3%, and 47.5%. GDPval-AA v2.1, a professional-work Elo across 44 occupations, is 1846 against 1735, 1708, 1542, and 1588. Anthropic still warned that at this level, “benchmark margins have become a less reliable guide to real-world differences,” and that in its own use the gap with Fable 5.1 is narrower than the scores suggest.

The cost comparisons the company wants circulating are configuration-specific. At default effort on FrontierCode, it said Opus 5.5 beats Astra at roughly 20% of the cost per task. On Terminal-Bench 4.0 it said the model matches Astra for about 40% of the cost. Those are not claims that every API call is 80% cheaper than Astra. Sol still lists at half of Opus 5.5 on fresh input and output, $2 and $10 against $4 and $20, so Opus has to save steps, retries, or reviewer time to win the invoice.

Customer tests are more useful than the trophy table because they measure finished work. GitHub’s early access ran the model in Copilot CLI and VS Code.

Developers want agents that can take on real software work and finish it. In our testing across GitHub Copilot CLI and VS Code, Claude Opus 5.5 used among the fewest tokens and steps we measured. In VS Code, it solved more terminal tasks than Opus 5 in less than half the steps. More than making individual tasks more efficient, it’s making developers’ bigger projects more achievable.

Mario Rodriguez, Chief Product Officer, GitHub

Clio staff developer Sean Heintz left it on a six-repo job overnight for more than 18 hours. Box said it used a third of the tokens Opus 5 did, with answers 40% less verbose. Hebbia said end-to-end finance workflows covered 86.6% of its rubric against 60.3% for Opus 5. Deloitte Consulting CIO Carl Bennett said even the lowest effort setting caught 72% of known bugs in code review against Opus 5’s 56% at high effort. Sonar’s evaluation found 42% fewer bugs, vulnerabilities, and code smells than Opus 5. An internal Anthropic rewrite of HAProxy from C to Rust finished in 9.5 hours against 12 for Fable 5.1 and cost 51% less, with both passing nearly all of HAProxy’s own tests.

Clearer writing is part of the cost story, because shorter, plainer answers are cheaper to generate and cheaper to check. In Anthropic’s side-by-side, Opus 5.5 explains a billing-period bug by putting the dollar split first, $1.50 from a free-tier change and $9.92 from a bad commit, then the broken midnight comparison. Opus 5 opened on the refactor and the interval math. Ramp staff engineer John Ruelas said verbose output had been his biggest frustration with frontier models, and that Opus 5.5 “writes like a good colleague.”

A Fallback Model Still Handles the Cyber Queue

Opus 5.5 is the first Opus to launch with a similar class of safeguards to Fable 5.1 in cybersecurity, biology, and anti-distillation. Anthropic said it is comparable to Claude Mythos 5.1 in biology and cybersecurity, so flagged requests fall back to another model: cybersecurity tasks to Claude Opus 4.8, and biology and frontier-LLM-development tasks to Opus 5. Those fallbacks likely lower some published scores, which the company disclosed. It also said the model posted its best result to date on an automated behavioral audit run across thousands of simulated scenarios, and that external testers including Frontier Design and METR saw it before release.

The Fable-class cyber safeguards on the new Opus sit on a classifier that screens each action, an open-source sandbox security teams can audit, and code review meant to catch vulnerabilities before merge. On a Gray Swan prompt-injection benchmark, Anthropic said Opus 5.5 ties Fable 5.1 for the lowest success rate of any model tested. Vetted labs can apply to a Life Sciences Verification Program for biology research, and a Cyber Verification Program is set to expand in the coming weeks so verified practitioners can use Opus 5.5 for that work.

The timing is awkward on purpose. The model is Anthropic’s first release since chief executive Dario Amodei argued that AI progress should be paced so safety work stays ahead of new skills, after lawmakers already asked frontier labs to write their own speed limit. Shipping a cheaper daily driver with Fable-class locks, then routing the sensitive slice to older models, is how Anthropic is trying to sell speed and restraint in the same week.

Anthropic Still Has to Sell This Before It Lists

The company has been lining up a public listing around this product cycle. It closed a $65 billion Series H on May 28, 2026, at a $965 billion valuation, and filed a confidential draft S-1 on June 1. The Series H materials put annualized run-rate revenue at $47 billion as of mid-May. Later figures put that run rate at $65 billion by the end of July, with more than 1,000 customers spending over $1 million a year. People involved in the process have described a November 2026 window rather than October, so third-quarter results can be shown, including a second straight quarter of positive adjusted operating income.

Ramp’s tracker of US business token spend has put GPT-6 Astra at about 13% against about 8% for Claude Fable, which is the momentum Anthropic is trying to blunt with a cheaper flagship rather than another $10/$50 model. Sonnet 5.5 and Haiku 5.5, the parts of the family that actually fill high-volume routes, are still “coming weeks” away. Until those land, Opus 5.5 has to do two jobs: keep coding agents on Claude, and give a listing story a unit-cost chart that is not only Fable.

THE DATES AROUND THE CUT

1. May 28, 2026: Closes a $65 billion Series H at a $965 billion valuation.
2. June 1, 2026: Files a confidential draft S-1.
3. July 24, 2026: Ships Opus 5 at $5 and $25, with no Opus list cut.
4. September 15, 2026: Mozilla puts the open-weight gap at 4.4 months.
5. September 21, 2026: Software firms are described as moving to open-weight models, including Chinese ones, on cost.
6. September 22, 2026: Claude Opus 5.5 lists at $4 and $20, with $0.20 cache reads.

The 20% list cut is small against Sol and smaller against self-hosted GLM or Kimi. The 60% cache-read cut is the piece that can change an overnight agent bill, which is why Anthropic leaned on typical-workload math instead of the sticker. Buyers who already left for open weights will not come back for 20% off input. Some of them will replay their own traces at medium effort, watch the cache hit rate, and decide whether 40% off a Claude job is enough to stay.

Frequently Asked Questions

What Is the Claude Opus 5.5 Model ID and Output Limit?

The API model ID is claude-opus-5-5. It uses a 1 million token context window, unchanged from Opus 5, and supports up to 128,000 synchronous output tokens, with image and text input.

Does Fast Mode Work on AWS, Google Cloud, or Microsoft Foundry?

No. Fast mode is a research preview on the first-party Claude API only, at $8 per million input tokens and $40 per million output tokens, and it is not sold on Claude Platform on AWS or on partner-operated clouds. It also cannot be combined with the Batch API.

How Much Does Batch Processing Cost on Opus 5.5?

Batch input is $2 per million tokens and batch output is $10 per million tokens, a 50% discount to the standard $4 and $20 rates, billed asynchronously for large request volumes.

What Does US-Only Inference Add to the Bill?

Setting inference to the United States applies a 1.1× multiplier to input, output, cache writes, and cache reads on Claude 4.6 and later models, including Opus 5.5. Global routing, the default, uses the standard sheet.

Can Thinking Be Turned Off on Opus 5.5?

No. Thinking is always on. Medium is the API default, and quality and cost are changed through the effort setting (low, medium, high, xhigh, max) rather than by disabling thinking.

When Does the Paid-Plan Usage Reset Expire?

Pro, Max, and Team users can apply the launch reset from Settings, Usage any time until October 22, 2026. Anthropic did not change seat prices; the extra room is an allowance and a cheaper token rate, not a cheaper subscription.

Disclaimer: This article is news reporting and analysis of a model launch, published prices, and related commercial context. It is informational only and is not investment advice, a solicitation, or a recommendation to buy or sell any security, including any future Anthropic shares. Readers who are weighing a financial decision should consult a licensed financial adviser who can review their situation. List prices, usage limits, cloud availability, benchmark scores, and any listing timetable reflect the sources used here and can change without notice.