TechNewsFirst

Anthropic put a number on its own price cut this week, and most of it checks out.

On September 22, 2026, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5 and $25, according to Anthropic's own pricing documentation. That is a flat 20% cut on both ends, no caveats attached. Cache reads dropped further, from $0.50 to $0.20 per million tokens, a 60% cut. Cache writes came down too: $5 per million for a 5-minute cache, $8 per million for a 1-hour cache. Batch processing is half of the standard rate, $2 and $10. Those are numbers a developer can drop straight into a cost model today, and they are the company's own published figures, not a marketing estimate.

The number Anthropic actually led with in coverage of the release is 40%. Multiple outlets reported Anthropic's framing that overall savings run "closer to 40%," because, in the company's own account, Opus 5.5 also needs fewer tokens to finish a given task and generates output more than 30% faster (Benzinga, Sept. 22, 2026; VentureBeat, Sept. 22, 2026). That second number is not a price. It is a claim about how the model behaves once you're using it, made by Anthropic about Anthropic's own model, with no independent benchmark published alongside it in this reporting cycle. It belongs in the estimate column, not the receipt column, and Anthropic's own technical documentation complicates it in a way worth sitting with.

Anthropic's Opus 5.5 release notes list several "behavior differences" from Opus 5 that show up without any code change. Two matter here. First: "the default effort is medium." A request that doesn't specify an effort level ran at high on Opus 5; on Opus 5.5, the same unmodified request now runs at medium, a shallower and cheaper setting, automatically. Second, and this is the one that undercuts the marketing framing directly: "at the same effort setting the model tends to think more per turn than Claude Opus 5, most of all at xhigh and max." That is Anthropic's own published language. It means that if a developer keeps the same effort setting they used on Opus 5, rather than accepting the new lower default, Opus 5.5 does not get cheaper. It gets more expensive, because it thinks longer at every fixed effort tier above the new default.

So the "closer to 40%" figure needs two separate things to be true at once: that you accept a real behavior change (less deliberation per request, by default) as free, and that your workload benefits from caching in roughly the proportions Anthropic modeled internally. Neither is guaranteed for any specific developer. An independent breakdown published around the launch (Digital Applied, Sept. 2026) put the realistic range closer to 29% for cache-heavy sessions and closer to 20%, meaning just the sticker-price cut and nothing more, for workloads that don't cache well. Forty percent is closer to a best case than a typical one.

There is a second way the sticker price and the real bill can diverge, and it comes from Anthropic's own safety architecture rather than from the pricing table. Opus 5.5 ships with safety classifiers for cybersecurity, biology, and frontier-LLM-development risk categories. When one fires, Anthropic's documentation confirms the request can be declined and retried through "server-side fallback," which reruns it "on the model Anthropic recommends for that category." Reporting on the release (The New Stack, Sept. 22, 2026) says cybersecurity-flagged requests typically fall back to Opus 4.8, while biology- and frontier-development-flagged requests fall back to Opus 5, the very model Opus 5.5 was just priced 20% below. A request billed under Opus 5.5's name can, in the exact cases Anthropic built a safety system to catch, actually run on an older model priced higher per token.

It gets worse for multi-turn agents specifically, and again this comes straight from Anthropic's own docs, not speculation. Every "thinking" block is tied to the model that produced it. Opus 5.5 can read thinking blocks from Opus 5 and earlier models, but the reverse often isn't true, and a request that gets routed to a different model mid-conversation, whether by a safety fallback, a cost router, or a retry after a timeout, has its unreadable thinking blocks silently dropped before the model ever sees them. The dropped blocks aren't billed. But the model handling that turn now has to redo the reasoning those blocks represented, from nothing, in a fresh set of output tokens that are billed. Anthropic's own migration guidance for this exact scenario is to keep conversations "append-only" and to treat model switches as something to actively guard against, which is an unusual thing to have to warn developers about in the same release notes that advertise a price cut.

None of this happened in isolation. Opus 5.5 launched about ninety minutes before OpenAI released GPT-6 Sol and Luna on the same day, cutting its own API prices roughly in half and undercutting Opus 5.5 outright: Sol's $2 input, $10 output per million tokens beats Opus 5.5 on both numbers (Yahoo Finance, Sept. 22, 2026; The Next Web, Sept. 22, 2026). xAI's Grok 4.7 had already set a $2 input-token bar the day before. Anthropic did not cut its price into a vacuum; it cut its price into an active three-way war on the same metric, on the same week.

And that week has a calendar on it. The Wall Street Journal reported on September 18, 2026, that Anthropic is targeting a November IPO at a valuation near $2 trillion, pushed back from an earlier October target specifically so the company can present a full quarter of third-quarter revenue to investors before pricing the offering (reporting relayed by PYMNTS and Cryptobriefing, Sept. 18-21, 2026). Anthropic's revenue run rate reportedly grew from $9 billion at the end of 2025 to $65 billion by the end of July 2026. A flagship model that benchmarks near the pricier Fable 5.1 tier, runs faster, and comes with a real, checkable price cut is exactly the release that makes a growth-and-efficiency story land better on a roadshow slide, three weeks before that roadshow is scheduled to start.

None of that requires the 20% price cut to be fake. It isn't; it's on Anthropic's own price list and it applies with no strings attached. What the calendar and the documentation together show is that the "closer to 40%" figure sits at the intersection of a real, unconditional cut, an unaudited efficiency estimate about how the model will behave for you specifically, and a business timeline with an unusually large number attached to it.

For anyone actually running this in production, the practical advice is the boring kind. Bank the 20% and the 60% cache-read cut; those are contractual. Don't assume the other 20 points without testing your own workload against your own effort setting, because Anthropic's own instructions are to "re-run your effort sweep" rather than carry a setting over, precisely because the model doesn't behave the same way at the same nominal effort level it used to. And if a pipeline uses server-side fallback, retries, or any multi-model routing, check which model actually served each turn before trusting the invoice. The bill will say Opus 5.5. The model that did the work, and the price attached to it, might not.

Cheaper frontier AI is a real trend this week, not a myth: Anthropic cut its base rate 20%, and OpenAI beat it by half an hour. But "40% cheaper" was never a price. It was Anthropic's own estimate of how a customer will use the model, published the same week the company began building the growth story it plans to sell to public investors.