Anthropic Trims Claude Opus 5.5 Pricing 20% Weeks Ahead of N

Anthropic has released Claude Opus 5.5, pricing its new flagship at $4 per million input tokens and $20 per million output tokens, a 20% reduction from the $5 and $25 charged for Opus 5. The model is available now through the Anthropic API, AWS, Google Cloud and the Microsoft ecosystem, and arrives weeks before the company's planned Nasdaq listing. The Claude Opus 5.5 pricing change makes it the first Opus generation to cost less than the model it replaces.

Anthropic frames the release as a token-efficiency upgrade, not a capability leap. The company says typical workloads run about 40% cheaper end to end because the model finishes tasks with fewer tokens, on top of the lower per-token rates. Default output speed improves by more than 30%, according to Anthropic, and the company is raising five-hour usage limits for Pro, Max and Team subscription plans while adding a savable rate-limit reset.

The gap between the 20% Claude Opus 5.5 pricing cut and the 40% cost-per-task claim matters for anyone building a budget. The first number is contractual and auditable. The second depends on workload shape, prompt design and how aggressively the model spends its thinking budget, none of which Anthropic can guarantee across a customer's traffic.

Claude Opus 5.5 Pricing: What Buyers Get

Opus 5.5 keeps the 1 million token context window of the Claude 5.5 family and supports up to 128,000 synchronous output tokens, with a 300,000-token batch mode in beta. Cache reads fall to $0.20 per million tokens, 60% below Opus 5, while cache writes cost $5 per million. A fast mode available in Claude Code and the Claude Platform reaches up to 2.5 times standard speed at $8 per million input tokens and $40 per million output tokens.

Metric	Opus 5	Opus 5.5
Input tokens (per million)	$5	$4
Output tokens (per million)	$25	$20
Cache reads (per million)	$0.50	$0.20

On the benchmarks Anthropic published, the model scores 66.4% on Terminal-Bench 4.0, 54.4% on Frontier Code v1.1, 57.8% on Cursor-Bench 4.0 and 1,846 Elo on GDPval-AA v2.1. Anthropic says those results beat Mythos 5.1 and Fable 5.1 on agentic coding and reasoning, and that Opus 5.5 matches Fable-level performance on most work.

That last claim carries the most weight for buyers. Fable 5.1 occupied a higher, more expensive tier, and Anthropic is now telling customers they can reach comparable results on agentic coding without paying for it. The premium tier loses much of its justification if the cheaper model holds its position in production.

For teams already on Opus 5, the move is a configuration change rather than a rewrite. Model identifiers update, prompt caches rebuild once, and the lower cache-read rate compounds across long agent sessions. The harder decision belongs to buyers who were paying for Fable 5.1 or Mythos 5.1, because staying on a premium tier now needs a justification that published benchmarks no longer supply.

The Catch Inside the Discount

Savings depend on workload shape, and Anthropic's configuration adds variables that are hard to model from the outside. Adaptive thinking is enabled by default with adjustable effort levels, so a team that leaves the setting untouched may spend more tokens than one that tunes it. Pushing effort to maximum can consume the entire 128,000-token output ceiling on a single request. Developer Simon Willison reported that a max-effort run on one SVG rendering task exhausted the output cap and cost $2.56 for an attempt that produced nothing usable.

Safety routing creates a second source of unpredictability. Requests flagged by classifiers for cybersecurity or biology topics can be silently sent to an older model, part of a safety update that also adds containment evaluations. For regulated industries that must document which system processed a given piece of data, silent rerouting is a governance problem instead of a line item. A developer calling Opus 5.5 is not always reaching Opus 5.5.

Cache economics cut the other way for agentic workloads. Coding agents replay large context windows on every step, so cached input usually dominates the token mix. Dropping cache reads from $0.50 to $0.20 per million tokens reduces that dominant component by 60%, which can produce savings well beyond the headline 20% for a well-cached agent loop. The 40% figure Anthropic cites is plausible in exactly that scenario and far less so for short, uncached, single-turn calls.

OpenAI Answers Within Hours

OpenAI released GPT-6 Sol and GPT-6 Luna less than two hours after Anthropic's launch, cutting prices roughly in half. Sol carries a 90% discount on cached input, which puts its cache reads at $0.20 per million tokens, the same rate Anthropic charges. GPT-6 Astra, OpenAI's top tier, holds at $10 per million input tokens and $50 per million output, leaving room above the new floor. xAI shipped Grok 4.7 inside the same 48-hour window.

Both releases target agentic coding, where a single task can trigger dozens of model calls and the token bill scales with loop length instead of conversation length. That is why cache pricing moved faster than list pricing on both sides of the market. A tool that runs an agent for twenty minutes cares far more about the cost of re-reading its own context than about the price of one fresh prompt. Anthropic's own benchmark selection points the same way, since Terminal-Bench, Cursor-Bench and Frontier Code all measure multi-step task completion.

For buyers, the effect is a repriced market rather than a single cheaper option. A price point that looked aggressive on Tuesday morning was matched by the afternoon, which shortens the useful life of any inference budget modeled on list rates. Multi-model routing, already common in agent frameworks, shifts from optimization to default cost control.

The IPO in the Background

Anthropic has filed confidentially for a Nasdaq listing at a $965 billion valuation, with a debut expected in the coming weeks. That timeline reframes what a model launch is for. Public-market investors weigh gross margin, net revenue retention and pricing power, and inference cost is one of the largest components of cost of goods sold for a model vendor. A flagship that costs 20% less per token while claiming 40% lower cost per task supports the argument that Anthropic can grow enterprise revenue without letting margin erode.

The cost floor reaches past list prices. Anthropic is also raising five-hour limits on Pro, Max and Team plans, which moves some consumption out of metered API billing and into flat-rate subscriptions. That mix protects retention among individual developers while the enterprise side absorbs the per-token cuts.

The timing also collides with the safety rhetoric that preceded it. Dario Amodei and Sam Altman both publicly urged the industry to slow frontier development ten days before these launches. Opus 5.5 and the GPT-6 pair are the first releases from either company since those statements. Neither proposal called for an immediate halt, and both new model families attack cost instead of capability, which narrows the apparent contradiction between what the executives said and what the companies shipped.

Competitive pressure explains more of the day than any change in safety philosophy. Anthropic shipped Opus 5 only two months before cutting its successor's price, and enterprise buyers had been moving workloads to cheaper models. A vendor preparing for a public listing cannot afford to defend premium pricing while customers defect to lower-cost alternatives.

Why This Matters

Model competition has moved from benchmark leadership to cost per completed task, and the same-day responses from both major vendors confirm that the shift is structural. Enterprise buyers gain leverage: cheaper flagships from Anthropic and OpenAI within hours of each other make multi-vendor routing the sensible default, and per-task measurement more useful than per-token list prices. For Anthropic, the cut doubles as an IPO argument about margin discipline. The next signal to watch is whether the $4/$20 rate holds through the quarter, or gets cut again before the listing.

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.