Anthropic Cuts Opus 5.5 Costs 40% Before a $2 Trillion Bet

Anthropic launched Claude Opus 5.5 on September 22, and says typical jobs now costs 40% less to run than Opus 5. It is the first model in the Claude 5.5 family and, the company says, matches Claude Fable 5.1 on most work. List prices fell 20% to $4 per million input tokens and $20 per million output tokens. The larger cut sits on cache reads, the line that dominates long agent sessions.

OpenAI shipped GPT-6 Sol and GPT-6 Luna later that same Tuesday and halved those API rates. Ten days earlier, Anthropic chief executive Dario Amodei had asked labs to slow capability gains. Bankers working a possible November listing have already told investors the Claude maker could be valued near $2 trillion. The cheaper model is the product Anthropic is putting in front of that pitch.

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

The Cache Cut That Makes Overnight Agents Cheaper

Anthropic’s 40% figure is not the list-price cut. Input and output each dropped 20% from Opus 5’s $5 and $25. Cache writes fell to $5 from $6.25. The company says Opus 5.5 also uses fewer tokens per task and generates output more than 30% faster, and that those pieces together produce the 40% saving at default settings.

The line agent builders actually price against is cache. Anthropic now bills cache reads at $0.20 per million tokens, 60% below Opus 5’s $0.50, and says those reads make up most of the cost of agentic and coding work. Fast mode in Claude Code and on the Claude Platform runs up to 2.5x quicker at $8 input and $40 output. US-only inference is offered at 1.1x on input and output. The API name is claude-opus-5-5.

API PRICES PER MILLION TOKENS

Pro, Max, Team, and seat-based Enterprise plans also get higher five-hour usage limits and a rate-limit reset that subscribers can save and spend later. Pro is $17 a month on annual billing, or $20 billed monthly. Max starts at $100 a month. That does not change the API math that will show up in enterprise invoices.

OpenAI Halved GPT-6 Sol the Same Afternoon

OpenAI’s post said improvements in caching and inference let it cut GPT-6 Sol and Luna API prices 50% from GPT-5.6 promotional rates, and that the new rates are permanent. Sol is $2 input and $10 output, half of GPT-5.6 Sol’s $4 and $20. Luna is $0.10 and $0.50. Cached Sol input is $0.20, matching Opus 5.5 on that one line and undercutting it on everything else.

On raw list prices, Sol costs half what Opus 5.5 costs for fresh input and output. That is the comparison procurement teams will print. Anthropic’s 40% saving is measured against its own last Opus, not against Tuesday’s OpenAI floor. Independent testers keep making the same split: Opus 5.5 scores higher on hard coding jobs, Sol is cheaper per task until you spend up the effort ladder.

Artificial Analysis’s independent scores for Sol and Opus 5.5 put GPT-6 Sol (max) at 48 on its Intelligence Index against 58 for Opus 5.5 at max effort, with Sol at $1.06 a task and Opus 5.5 at $5.98 on that harness. Those are not Anthropic’s numbers. They are the closest public side-by-side run of the two models that shipped on the same day.

A 680,000-Line Migration Finished in Under a Day

The launch’s working examples are long jobs, not chatbot answers. Anthropic says Opus 5.5 is built for migrations, audits, and other work that used to burn a team for weeks. The specimens it published are specific.

EARLY JOBS ANTHROPIC CITES

• Code migration: One tester finished a 680,000-line migration in less than a day, work the company says would have taken an engineering team weeks.
• Audit and repair: Another tester audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took more than 20 hours and used 2.5 times as many tokens.
• C-to-Rust rewrite: In an internal test translating HAProxy from C to Rust, both Opus 5.5 and Fable 5.1 passed nearly all of HAProxy’s regression tests, but Opus 5.5 finished in 9.5 hours against 12 and cost 51% less.

Asked to cut load times across every page of a web app, Opus 5.5 succeeded 39 of 40 times, while Opus 5 made smaller changes that also altered the app’s behavior. Anthropic itself warns that at this level, benchmark gaps are a weaker guide to lived difference, and that the gap with Fable 5.1 is narrower in daily use than the table implies.

What GitHub Copilot Measured in VS Code

Named customers on the launch page described fewer steps, not just a lower sticker. Quantium’s Harley Barnes, executive manager for AI technology, said a coding task that had taken 38 prompts over four days came in at 11 prompts over three hours. Clio staff software developer Sean Heintz said he handed the model a job across six repositories and let it run unattended for more than 18 hours. Optiver’s global head of AI engineering, Noyan Tokgozoglu, said agentic coding tasks matched Opus 5 quality in about half the turns, time, and output tokens, cutting that workload’s cost 40 to 50%.

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

Dan Shipper, who has been running the model on coding, design, and knowledge work, said it is pulling some Codex users back to Claude on large end-to-end builds and has become a daily pick for ambitious product work. He also said it still trails OpenAI’s top models on writing, and that it has a habit of doing too much, including at least one knowledge-work test that ran past a 10-minute limit. That is the trade buyers are actually arguing about: cheaper long jobs, with a model that sometimes will not stop.

Opus 5.5 Leads Terminal-Bench and Trails on Science

Anthropic’s own table, run with production safeguards on and adaptive thinking at max effort unless noted, puts Opus 5.5 ahead of Fable 5.1, Opus 5, and GPT-6 Astra on several coding and knowledge-work tests. Terminal-Bench 4.0 is reported at xhigh effort for Opus 5.5, with a standard error of ±2.6 points. The public board did not yet list Opus 5.5 on launch day, so treat 66.4% as Anthropic’s harness, not a third-party leaderboard print.

ANTHROPIC’S LAUNCH BENCHMARK TABLE

Astra still leads AutomationBench at 41.4% against 40.0%, and Terminal-Bench-Science at 64.6% against 58.7%. When Anthropic’s safeguards stepped in, cybersecurity tasks fell back to Claude Opus 4.8 and biology tasks to Opus 5, which the company says likely reduced some scores. CursorBench 4.0, a long-horizon editor-agent set published by Cursor on September 10, has Opus 5.5 at 57.8% against 41.7% for GPT-5.6 Sol. GPT-6 Sol is not in that Anthropic column.

Ten Days After a Call to Pace the Frontier

Opus 5.5 is Anthropic’s first release since Amodei called for pacing the frontier in a 3,800-word essay on September 12. He wrote that AI helping to build the next AI had sped up since summer, and that safety work now needed time more than slogans. OpenAI chief executive Sam Altman and Elon Musk both said they agreed with that slower tempo. Musk then replied “Congrats!” to Claude’s launch post on Tuesday.

We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.

Dario Amodei, CEO, We Must Pace the Frontier

The new model was tested before release by external evaluators including METR and Frontier Design. Anthropic says it posts the strongest score to date on the firm’s automated behavioral audit, is more resistant than Opus 5 to prompt injection, and is less likely to take hard-to-reverse actions. It ships with Fable-class safeguards on biology, cybersecurity, and anti-distillation. Vetted labs can apply to a Life Sciences Verification Program; a Cyber Verification Program is due to widen in the coming weeks.

US President Donald Trump took the other side the same day, in remarks at the United Nations General Assembly. “I’m not going to stifle growth of something that will be bigger than the Industrial Revolution,” he said, adding that the Justice Department would watch the sector. He rejected what he called a “globalist scheme” to control the technology and said US documents should use “super intelligence” instead of “artificial intelligence.”

THE FORTNIGHT AROUND THE LAUNCH

1. May 28, 2026: Anthropic closes a $65 billion Series H at a $965 billion valuation.
2. June 1, 2026: The company confidentially files a draft S-1 with the SEC.
3. September 12, 2026: Amodei publishes We Must Pace the Frontier.
4. September 22, 2026: Claude Opus 5.5 and GPT-6 Sol and Luna ship; Trump tells the UN he will not stifle growth.
5. November 2026: Target window for a listing, with a roadshow expected in mid-October, before the November 3 midterms.

Pacing, in Amodei’s own wording, does not mean stopping training. It means taking time to align models and letting third parties check that work. Shipping a faster, cheaper Opus 10 days later is how that sentence looks in a market where OpenAI is cutting list prices in half on the same calendar day.

A $2 Trillion Listing Still Has No Final Terms

The commercial clock is the listing. Anthropic has not confirmed a valuation, a raise, or a date. Bankers on the offering have told potential investors the company could seek more than $100 billion at about $2 trillion, which would top SpaceX’s June debut at $1.77 trillion and $85.7 billion raised. In May, Anthropic’s annualized revenue had passed $47 billion. Investors modeling the float have put year-end annualized revenue between $100 billion and $120 billion. Those are investor figures, not a company target.

WHAT WE KNOW

• Last private mark: Series H priced Anthropic at $965 billion on May 28, 2026, after a $65 billion round.
• The filing: A confidential draft S-1 went to the SEC on June 1, 2026.
• The product: Opus 5.5 is live on Claude for Pro, Max, Team, and Enterprise users, and on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Foundry.

WHAT IS UNCONFIRMED

• Valuation and raise: The $2 trillion figure and a raise above $100 billion come from banker conversations with potential investors, not from Anthropic.
• Timing: November is the working window so third-quarter results can be shown; the prospectus is not yet public.
• The rest of the family: Claude Sonnet 5.5 and Claude Haiku 5.5 are promised in the coming weeks, with no dates.

A 40% cheaper Opus helps the volume story that a $2 trillion pitch needs. It does not, on its face, help the margin story, and Sol’s $2 and $10 list is the number sitting on the other side of every enterprise bake-off. Hyperscalers already carry the model. The next cheaper SKUs, Sonnet 5.5 and Haiku 5.5, are how Anthropic would fill the rest of the invoice. Until those ships, Opus 5.5 is the bet: long jobs, lower unit cost than last quarter, still twice Sol on the open token price, and a public listing that still has no final terms.

Disclaimer: This article is news reporting and analysis of a product launch and of public comments about a possible share listing. It is for information only and is not investment advice, a solicitation to buy or sell any security, or a recommendation of any AI product. Readers considering an investment in Anthropic, OpenAI, cloud providers, or related companies should consult a licensed financial adviser and review official filings if and when they appear. Prices, benchmarks, revenue run rates, and listing terms are those given by the named companies, bankers, or investors as of the dates cited and can change.