Anthropic Cut Opus 40%. Its own flagship was down to 3.7% of spend.

On September 22 Anthropic shipped Claude Opus 5.5 with one sentence: "It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5."

Read as a launch, that is a competitor story. Everyone read it that way, because Reuters had reported three days earlier that Anthropic was considering a model to counter GPT-6 Astra's momentum.

The more interesting reading arrived the next morning, from Tomasz Tunguz, and it points inward rather than outward.

Four releases at one price, then a cut. And the reason is not Astra.

The cut has a cause, and it is on Anthropic's own shelf

Tunguz supplied the demand-side numbers in the same piece, and they are the part worth carrying into a planning conversation.

Claude Fable 5.1 took 3.7% of gateway spending in its first twelve days. Fable 5 before it had already fallen from a 13.2% peak to 4.9%, and Tunguz names when: "when Claude Opus 5 shipped at half the price."

So the sequence reads differently. Anthropic put out a flagship at $10 and $50. Then it put out Opus 5 at half that, and Opus took the flagship's share. Then in September it cut Opus too.

This is not primarily a counter-launch. It is a company finishing the substitution of its own expensive model. Astra is the narrative; the internal cannibalisation is the mechanism. And the part nobody said on launch day is that the Anthropic flagship was never the expensive option. Fable was.

What the vendor claimed, and what the one operator with a week of use said

Dan Shipper's team at Every had pre-release access and ran it across coding, design, writing and knowledge work.

Put that beside Anthropic's sentence and the gap is visible. Anthropic says performs at the level of Fable 5.1 for most tasks. Shipper says sometimes beats it. Anthropic says it costs 40% less. Shipper says up to 40% cheaper.

The independent operator is weaker than the vendor on both axes, which is the normal direction and worth saying out loud, because the number that will travel is Anthropic's. As of September 29 no independent benchmark has scored Opus 5.5 against Fable 5.1. Not Artificial Analysis, not Ramp's accounting bench, not AutomationBench. Every claim in circulation is first-party or vibe.

The switching signal underneath it

Shipper's most useful sentence is not the comparison. It is that Opus 5.5 is "pulling some of our recent Codex converts back to Claude."

Three months earlier, the same team went the other way: Codex captured nearly all of Shipper's coding usage when Anthropic disabled Fable in June. Same house, same operator, flow reversed in a quarter.

Peter Yang, who defected in the same window, put the comeback framing on it: "With Opus 5.5, I think Claude is finally back." The premise inside that sentence is the real finding. Anthropic had been widely read as behind since the Astra launch, and this is the first item that runs the other way.

Read both together and the planning conclusion is uncomfortable for anyone standardising: harness loyalty at the frontier is measured in weeks, not years. Any client architecture that assumes a model choice holds for a year is wrong on the observed evidence. Build on the context, skills and workflows the client owns, and keep the model swappable. See GPT-6 Astra and Claude Fable Now Cost the Same.

Two operating rules you can hand a team this week

The most transplantable thing in the whole launch came from Every's practitioner notes, and neither line is about capability:

Set a time goal. "It won't stop on its own, and it'll eat your weekly limit. Say '10 minutes, produce X.'"

Name the deliverable. Asked for a run of show, it produced training materials and handouts and missed the artifact that was actually requested.

Both are consequences of a more agentic model rather than a worse one. An agent with no natural stopping point is bounded by budget instead of by finishing, and a cheaper model that runs longer does not necessarily cost less per week. The mitigation is a sentence, not a system: state the artifact and the time box.

And then another operator said the opposite

Four days later, Signüll, running it daily:

Two first-party operators, four days apart, in opposite directions on the same axis. Nothing published reconciles them, and the boring explanations are the likely ones: different plans, different task mixes, and a possible change in quota accounting between pre-release and general availability.

But notice they may not be contradictory at all. "It won't stop on its own" is a claim about the agent's behaviour. "Dramatically less of my quota" is a claim about cost per unit of work. A cheaper model that runs longer produces either experience depending on whether somebody supplied a stop condition. On that reading Signüll is simply the operator who already time-boxes, which makes Every's advice the resolution rather than the loser. Do not drop the time box on the strength of one enthusiastic post.

Three labs, forty-eight hours, one message

The individual launch matters less than the week it landed in.

On September 21, xAI shipped Grok 4.7, which Musk described as "a strong combination of intelligence, speed and low cost." On the 22nd, Anthropic shipped Opus 5.5 at Fable-level quality for 40% less. The same day, OpenAI shipped GPT-6 Sol and Luna as "faster and more affordable models to support work at scale."

Not one of the three claimed a capability jump. All three claimed the same output for materially less money. That is a price war wearing launch clothes, and it arrived as three simultaneous vendor decisions rather than as a trend line somebody drew afterwards.

For a company planning AI spend, that is the fact of the month. Not that a model got better, but that the frontier now reprices monthly and the direction is down. Anything you are building that depends on a specific model at a specific price has a shelf life measured in weeks.

What to hold loosely

The $3 and $15 figures are derived here and published by nobody. Anthropic gave a percentage against an unstated base; Tunguz supplied the base; the arithmetic is ours. His own GPT-5.6 pricing in the same essay disagrees with figures elsewhere in our records, so treat the Opus series as plausible rather than confirmed.

Matthew Berman's launch-day line that Opus 5.5 is "the absolute best model on the planet" is still unsupported by any benchmark. His other half, that it is "cheaper than Astra/Fable," now looks true on price by roughly 3x, and it was already true of Opus 5 before the cut.

And the writing-quality claim never settled. Polymarket relayed that Anthropic "may have finally fixed AI's slop writing problem" on the 22nd, which is aggregated sentiment with no sample behind it. The next morning Steve Yegge, after months on Fable, said GPT-6 Astra's writing is "a delight." Different comparison, opposite mood, and no prose benchmark exists on either side.

What survives all of it is the pricing shape, because it is the only part with numbers behind it. Anthropic held one Opus price across four releases, cut it the first time its own flagship stopped selling, and did it in the same forty-eight hours two competitors cut theirs.