Claude Opus 5 Lands at Half of Fable 5's Price

Anthropic made Claude Opus 5 generally available on July 24, 2026, priced at $5 per million input tokens and $25 per million output, which the company's own materials put at half the rate of Fable 5. That single number is why the launch got the attention it did, and it did not come out of nowhere. The specs had already leaked, the positioning had already been telegraphed by the models Anthropic shipped ahead of it, and the price undercut was the last piece to drop.

From a Honeycomb config leak to a Friday GA

The sequence ran over roughly three weeks and it was unusually legible for an Anthropic release.

Earlier in July, a configuration reference under the internal name "Honeycomb" surfaced and got picked apart by people who track model plumbing. Coverage of that leak described a near-frontier model with a 1M-token context window and adjustable effort controls. At the time it read as a preview of an unannounced Opus. It turned out to be exactly that: the shipped Opus 5 carries the same 1M context and the same effort-dial behavior the leak pointed to.

Before Honeycomb, Anthropic had already been laying out a tiered stack in public. Claude Sonnet 5 shipped as the workhorse agentic tier. Fable 5, the Mythos-class flagship, came back to worldwide availability on July 1 once export restrictions lifted. Mythos 5 stayed positioned as the restricted, high-capability tier for cyber and bio work that Anthropic gates behind additional controls. So by mid-July there was a clear ladder with one obvious empty rung: a broadly available, high-capability model priced for production volume rather than for frontier bragging rights.

Opus 5 filled that rung on July 24. Anthropic's launch post puts numbers on a spread of third-party boards: Frontier-Bench v0.1, where it says Opus 5 surpasses every other model and more than doubles Opus 4.8; CursorBench 3.2, within 0.5% of Fable 5's peak at half the cost; ARC-AGI 3, three times the next-best model; Zapier's AutomationBench, around 1.5x the next-best pass rate; and OSWorld 2.0, past Fable 5's best result at just over a third of the cost. Two internal evals sit alongside them, organic chemistry up 10.2 percentage points on Opus 4.8 and protein tasks up 7.7. The same post names where Opus 5 still loses. It stays behind Mythos 5 on cybersecurity exploitation and on long-running autonomous research in biology. A vendor publishing its own ceiling is worth more than any of the boards above it. The launch drew immediate discussion on X, including a widely shared thread quoting Anthropic's own account. The benchmark figures in circulation are, for now, Anthropic's numbers on Anthropic's chosen tasks. Independent agentic leaderboards had not published a full Opus 5 line at launch, so the third-party read is still catching up to the vendor's.

The pricing is the part that is fully verifiable, because it is a published rate you can compare against the rest of the lineup:

Anthropic did not publish Fable 5's rate as a headline in the Opus 5 post; the "half the price" comparison is theirs, and I am reproducing it as a claim, not confirming the exact Fable 5 sheet independently.

The half-price cut points at Fable 5's floor, not Opus 5's ceiling

Read the ladder in order and a pattern falls out. Anthropic is not using Opus 5 to claim the top of any leaderboard. It is using Opus 5 to reset what "near-frontier, generally available" should cost, while keeping the genuinely top-end capability behind Fable 5 and behind the restricted Mythos 5 tier.

That is a deliberate segmentation, and it explains the effort controls better than the benchmark scores do. A dial that lets you trade reasoning depth for speed and cost only matters if you expect people to run the model at high volume on tasks that do not all need maximum effort. Frontier-flexing models do not ship cost dials; production models do. The 1M context and the Zapier and OSWorld agent evals point the same way: this is a model built to sit inside agents and workflows that make a lot of calls, where a 2× price gap compounds fast.

If that reading is right, the next move is competitive pressure downward on the whole near-frontier band. Anthropic has effectively told buyers that Fable 5's capability is worth a premium and Opus 5's capability is worth half of it. That is a bet that most production traffic does not need the flagship, which is probably true for the bulk of agent and coding workloads. It also puts a public number on the table that OpenAI's GPT-5.6 tier and xAI's Grok 4.5 now have to answer, because "roughly Opus-class at $5/$25" is a much sharper anchor than the fuzzy tier comparisons buyers were making a month ago.

The vulnerability in this whole picture is the benchmark half. The price is real and checkable. The performance claims are vendor-run until the independent boards weigh in. If Opus 5 turns out to sit meaningfully below the flagships on the tasks buyers actually care about, then "half price" is just a discount on a lesser model, not a repricing of near-frontier capability. The strategy only holds if the "near" in near-frontier is genuinely near.

Two moves that would break the read

The cleanest way to falsify all of the above is for a competitor to match Opus 5's benchmarks below its price. If an independent agentic leaderboard, Artificial Analysis being the obvious one to watch, publishes an Opus 5 line and a cheaper model lands at or above it on Frontier-Bench, CursorBench, or the OSWorld and Zapier agent evals, then Anthropic's price anchor stops being an anchor. The $5/$25 point only reshapes the market if nobody credible is sitting under it at the same capability. GPT-5.6 and Grok 4.5 both have pricing sheets that could move.

The other move that would break it comes from inside Anthropic. If Fable 5 gets a matching price cut in the coming weeks, the tiering thesis collapses, because the entire read here depends on Anthropic holding a real price gap between the flagship and the volume model. A Fable 5 that costs the same as Opus 5 would mean the company was not segmenting on price at all, just staging two launches. Watch the Anthropic pricing page and the next independent Opus 5 benchmark line. Whichever of those two lands first will tell you whether the half-price cut was a repricing of the field or just a cheaper seat in it.