This offensive comes as the group prepares a Nasdaq listing that could, according to the Financial Times, target a valuation above $2 trillion. That figure is not yet an established valuation: it will depend on IPO conditions and investor appetite.
Claude Opus 5.5 sharply cuts the bill
Two months after Claude Opus 5, already designed to make AI agents more autonomous and less expensive, Anthropic is taking economic optimization much further.
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Opus 5. The advertised price reduction therefore reaches 20%. Cache reads, however, fall to $0.20 per million tokens, or 60% less. Because agents and coding tasks reuse large amounts of cached context, Anthropic estimates that the actual cost of a typical workload ultimately falls by around 40%. The model also generates responses more than 30% faster.
The difference becomes even clearer compared with Fable 5.1, priced at $10 for input and $50 for output. Anthropic nevertheless claims that Opus 5.5 reaches a comparable level on most tasks.
This is no longer just a race to build the most powerful model. Running costs have themselves become a weapon.
Anthropic is gaining ground particularly in coding
The results published by Anthropic give an indication of the leap. On Terminal-Bench 4.0, which focuses on complex tasks carried out in a terminal, Opus 5.5 scores 66.4%, compared with 52.3% for Opus 5 and 55.8% for Fable 5.1. GPT-6 Astra is listed at 57.9% in Anthropic’s comparison.
On FrontierCode, Opus 5.5 reaches 54.4% in the published maximum results, compared with 48% for Opus 5. On GDPval-AA, which focuses on professional work across 44 occupations, it reaches 1,846 Elo, ahead of Fable 5.1 at 1,735.
The gains are particularly notable on long-running assignments. Anthropic reports that one tester used Opus 5.5 to audit and fix 200,000 lines of code in less than three hours, whereas Opus 5 took more than 20 hours and used 2.5 times as many tokens. In another test, a migration of HAProxy from C to Rust was completed in 9.5 hours, compared with 12 hours for Fable 5.1, at a 51% lower cost.
These capabilities arrive as Claude is already being used for far more advanced cybersecurity tasks. Anthropic is therefore also strengthening its safeguards in cybersecurity and biology.
The $2 trillion IPO changes the scale
The timing is difficult to ignore.
Anthropic is preparing a public offering that could become one of the largest in technology history. The Financial Times reports that the company is targeting Nasdaq and that a valuation above $2 trillion is being considered. Anthropic also reportedly told investors that it was profitable on an adjusted operating basis for a second consecutive quarter. Its annualized revenue reportedly reached around $65 billion in July, compared with $9 billion at the end of 2025.
At this level, releasing a more powerful model is no longer enough. Anthropic must show that it can serve millions of users and businesses without inference costs consuming its growth. This pressure is connected to the race to build AI data centers capable of powering major models.
Opus 5.5 addresses this problem directly.
Faster.
Less token-intensive.
Cheaper to serve.
And powerful enough to handle more autonomous tasks.
The release also comes after Dario Amodei called for slowing the deployment of certain cutting-edge capabilities in order to strengthen safety mechanisms. Anthropic describes Opus 5.5 as its first major release since that position was taken, with additional external evaluations carried out before publication.
For Anthropic, the next showdown with OpenAI will therefore no longer be fought solely on benchmarks. It will be fought over how much intelligence a company can buy for one dollar of compute.
And just before an IPO potentially valued in the trillions, that figure is becoming almost as important as the model’s raw power itself.
