On 24 July, Anthropic shipped Claude Opus 5 at 5 dollars per million input tokens and 25 dollars per million output tokens. That is the interesting number, and it has nothing to do with a benchmark. It is the same price Anthropic charged for Opus 4.8, and it is exactly half of what the company charges for Fable 5, the flagship it made widely available in June (Anthropic pricing). Fable 5 lists at 10 dollars in and 50 dollars out.
Read that as a corporate action rather than a product launch. A frontier lab has priced its newest model to undercut its own top SKU by 50 percent, roughly a month after putting that SKU on the price list. Nobody forced the discount. Anthropic did it to itself.
The move, not the leaderboard
The launch materials lead with agentic benchmarks, and they are worth stating plainly as what they are: vendor-reported. Anthropic says Opus 5 beats every other model at any given cost on OSWorld 2.0, clears Fable 5’s best OSWorld result at about a third of the cost, scores three times the next-best model on ARC-AGI-3, and passes Zapier’s AutomationBench at around 1.5 times the next-best rate for the same cost per task. Those are the company’s own measurements, not independent ones, and the honest reading is to treat them as claims until third parties reproduce them. Anthropic also concedes, in the same post, that Opus 5 still trails its limited-availability Mythos 5 model on cybersecurity work, specifically on exploiting vulnerabilities. That is a factual limitation, not a scandal.
Set the scores aside, because the scores are not the story. The story is that near-frontier capability now costs 5 and 25 dollars, and the lab that built it chose to price the gap between “near-frontier” and “flagship” at zero incremental dollars over its own year-old Opus tier.
Capability became a dial
The other half of the launch is a parameter. Opus 5 exposes an effort control with tiers Anthropic labels high, xhigh, and max, letting a caller decide how hard a single request should think before it answers. Pair that with fast mode, a research-preview setting that runs Opus 5 at roughly 2.5 times the default output speed for twice the price, 10 dollars in and 50 dollars out, matching Fable 5’s list rate.
Put those two knobs together and the purchasing question changes shape. For most of the last three years the enterprise decision was a procurement decision: which model do I buy, and how much does the smart one cost versus the cheap one. Opus 5 reframes it as a runtime decision: given one model, how much effort should this particular request spend. Capability stops being a line item you select at contract time and becomes a parameter you set per call, sometimes per token budget.
That is a better deal for the buyer and a worse position for the seller, which is the entire point.
This is the July thesis, from the supply side
Three weeks ago we argued that the price of using a fixed level of machine intelligence was collapsing faster than almost any input cost in computing history, and that the labs building that intelligence would struggle to capture the value they created. That piece read the collapse from the outside, through list prices falling across vendors and generations.
Opus 5 is the same thesis observed from inside a single company, on a single day, by the company’s own hand. Anthropic did not wait for a competitor to undercut Fable 5. It undercut Fable 5 itself, and it did so while telling customers the cheaper model is good enough for most of what they were paying flagship rates to do. When a lab voluntarily compresses its own price umbrella one month after raising it, the umbrella was never structurally defensible. The value is real. The capture is the problem.
The ladder makes the compression visible. Anthropic’s current output prices run from Fable 5 and Mythos 5 at 50 dollars per million tokens, down through Opus 5 and Opus 4.8 at 25, Sonnet 5 at 10 during its introductory window, to Haiku 4.5 at 5 (pricing). Opus 5 lands the strongest non-flagship tier at the exact midpoint of the flagship, and holds it there rather than charging a premium for being newer.
One caveat keeps the comparison honest. Anthropic’s 4.7-era and later models, Opus 5 and Fable 5 included, use a tokenizer that emits roughly 30 percent more tokens for the same text than 4.6 and earlier. Between Opus 5 and Fable 5 that washes out, since both use the new tokenizer, so the 50 percent gap is real. Only when you compare across the tokenizer boundary does the effective price per unit of English drift from the sticker.
What it means
For anyone building on these models, the practical takeaways are narrow and concrete. First, the default reach for a flagship SKU is now hard to justify on capability grounds alone. Opus 5 covers most agentic and coding workloads at half the token cost, and the effort dial lets you spend flagship-grade compute only on the requests that need it. Second, the unit that matters for cost control is shifting from model selection to per-request effort budgeting, which is a different engineering discipline and a different line in the finance spreadsheet. Third, the 1M-token context window ships at standard rates across the full window, so long-context work no longer carries its own price tier.
For the labs, the read is starker. Anthropic has now demonstrated, twice in two months, that it will price its second-best model at half its best one and route customers there in its own launch copy. That is rational when your competitor will do it to you anyway. It is also an admission that the premium tier is a temporary perch, not a moat. The company that is best at making intelligence keeps proving it cannot hold the price of intelligence, and this time it proved it against itself.
The benchmarks will get argued over. The 5-and-25 price line will not. It is the same everywhere, it is half the flagship, and it is the clearest signal yet that in this market the maker captures less of the value than the math says it should.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.