# Anthropic Prices Opus 5.5 at $4 per Million Input Tokens, 20 Percent Below Opus 5

Anthropic says the model finishes the same work with fewer tokens, putting total cost per task about 40 percent below its predecessor while Meta sells frontier-tier tokens at less than a third of that list price.

- Published: 2026-09-25T06:22:14.083Z
- Canonical: https://polylog.news/ai/2026-09-25/anthropic-prices-opus-5-5-at-4-per-million-input-tokens-20-p
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic News](https://www.anthropic.com/claude-opus-5-5), [Polylog editors](https://polylog.news), [Anthropic News](https://www.anthropic.com/claude-fable-and-mythos-5-1)

Anthropic's [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) is the latest entrant in the contest that increasingly decides frontier application programming interface (API) market share: how much capability a vendor can deliver per dollar of inference. The company lists the model at 4 dollars per million input tokens and 20 dollars per million output tokens, down from 5 and 25 for Opus 5. Cache reads fell further, from 0.50 to 0.20 dollars per million tokens, [according to coverage of the release](https://venturebeat.com/technology/anthropic-releases-claude-opus-5-5-beating-fable-5-1-on-key-agentic-benchmarks-at-60-cheaper-api-price).

The more consequential claim is not the listed price. Anthropic says Opus 5.5 performs at the level of [Claude Fable 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1), its September flagship for coding and knowledge work, while costing about 40 percent less to run than Opus 5. That figure combines the roughly 20 percent price cut with the model spending fewer tokens to complete the same task. For anyone paying for long agentic runs, token count is the larger cost driver, and a model that reasons in fewer steps multiplies that saving across every call.

On Anthropic's own table, Opus 5.5 leads on SWE-bench Pro at 89.9 percent and Terminal-Bench 4.0 at 66.4 percent. Those are vendor-published numbers. The independent check comes from Artificial Analysis, which placed Opus 5.5 at the top of its Intelligence Index with a score of 58 at maximum effort, leading on six of ten core evaluations including Humanity's Last Exam at 61.4 percent and SciCode at 66.9 percent. When independent indices and vendor tables agree on the direction of a result, that is stronger evidence than either alone, though neither measures the failure modes that appear in production agent loops.

The competitive pressure affects both Anthropic and its rivals. The Telegram channel AI Post, which tracks lab releases, [reported that Opus 5.5 is pressuring OpenAI's GPT-6 Astra](https://t.me/aipost/8256) and speculated about a GPT-6.5 response. No lab has confirmed such a model, and that detail should be read as channel speculation rather than fact. What is documented is the price floor beneath the whole market: Meta charges [1.25 dollars input and 4.25 dollars output per million tokens](https://openrouter.ai/meta/muse-spark-1.3) for Muse Spark, and offers a 0.10 dollar tier to developers who let Meta train on their traffic.

## What this means

Capability parity at a lower price transfers margin from model vendors to the companies building on top of them. Anthropic is defending the high-value agentic coding segment, where customers tolerate premium pricing, by cutting cost per task rather than headline price alone, which protects gross margin better than a straight discount. The exposed parties are second-tier API vendors with neither a capability lead nor a cost structure that supports Meta-level pricing, and the beneficiaries are application companies whose unit economics improve without a code change. For Nvidia and the serving stack, fewer tokens per completed task cuts revenue per query but raises the number of tasks enterprises can afford to run.

## What to watch

- Whether OpenAI answers with a price cut or a new model tier rather than a capability claim, which would show that pricing, not benchmarks, is now the live competitive lever.
- Independent reproductions of the fewer-tokens-per-task claim on customer workloads, since the 40 percent saving depends on behavior that vendor evaluations measure under favorable conditions.
- Whether Meta's low-cost tiers pull volume away from premium models, which would tell you how much of the API market treats frontier capability as optional.