Anthropic's price ladder just did something that isn't supposed to happen. On June 9, 2026 it shipped Claude Fable 5 at $10 per million input tokens and $50 per million output — a new "Mythos-class" tier positioned above Opus 4.8, and priced at exactly double it. Six weeks later, on July 24, it shipped Claude Opus 5 at $5 per million input and $25 per million output. That is half of Fable 5. And on several benchmarks, the cheaper, newer model matches or beats the pricier one it launched beneath.
The reflex reading is that Anthropic cut prices, or that Fable 5 was overpriced. Both miss it. The more useful reading is that a frontier lab has stopped selling raw capability as its scarce good and started selling fit — pricing two strong models at two different points not because one is worse, but because they are aimed at two different buyers doing two different jobs. Do the arithmetic and the strategy shows through.
The numbers that don't line up the usual way
Start with the ladder as it stood after Opus 4.8 (May 28, 2026). Fable 5 is described in the facts as "exactly double" Opus 4.8, which pins Opus 4.8 at $5 input / $25 output. Fable 5 then landed above that at $10/$50, a genuine premium tier with a 1,000,000-token context window and, per third-party coverage, up to 128,000 tokens per response. Everything about it read as "the new top of the range costs more," which is how model ladders normally work.
Then Opus 5 came in underneath at $5/$25 — the same sticker as Opus 4.8, and half of the six-week-old flagship. Here is the anomaly laid out:
| Model | Released | Input / M | Output / M | Position |
|---|---|---|---|---|
| Opus 4.8 | May 28, 2026 | $5 | $25 | prior flagship |
| Fable 5 (Mythos-class) | Jun 9, 2026 | $10 | $50 | premium tier, above Opus 4.8 |
| Opus 5 | Jul 24, 2026 | $5 | $25 | default tier, below Fable 5 |
The newest strong model is the cheapest of the three. That inverts the "newer and better costs more" logic that has governed model pricing since GPT-3.
And the sticker understates it. Opus 5 ships with what Anthropic calls "adaptive thinking" — configurable reasoning effort, defaulting to high on the API and in Claude Code. A customer report puts Opus 5 at roughly one-seventh the reasoning tokens and half the latency of Opus 4.8. Reasoning tokens bill at the output rate, so token count drives cost as directly as price does. Take a reasoning-heavy task where Opus 4.8 emits 7,000 tokens of thinking and answer: that's 7,000 × $25/M = $0.175. If Opus 5 does the same job in ~1,000 tokens at the same $25/M rate, it's $0.025 — about seven times cheaper per task, from token efficiency alone, before you count the fact that it's also half of Fable 5's sticker.
Two caveats keep this honest. That one-seventh figure is a single customer report, not a controlled Anthropic number, so treat it as directional. And it's measured against Opus 4.8, not Fable 5 — we don't have Fable 5's token efficiency, so I won't claim a clean per-task multiple against the flagship. What we can say precisely: Opus 5 costs half of Fable 5 on sticker, and its effective per-task cost on reasoning-heavy work is lower still. (My analysis.)
Two prices because two buyers
The inversion stops looking strange the moment you stop treating the two models as points on one quality axis. They aren't competing for the same job.
Fable 5 is built for long-horizon autonomy — extended agentic coding over "days," per Anthropic's positioning, with minimal human intervention. Its third-party benchmarks are where it earns the premium: SWE-Bench Pro at 80.3% against Opus 4.8's 69.2% (and well ahead of GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%), SWE-bench Verified at 95.0%. The $10/$50 price buys reliability across a narrow set of workloads — hard, long-running, unsupervised agent work — where an incremental capability edge compounds over hours and is worth paying almost anything for.
Opus 5 is priced to be the default. Anthropic makes it the default model on Claude Max and the strongest available on Pro. Its wins cluster on the tasks the vast middle of usage actually consists of. On third-party CursorBench 3.2 it lands within 0.5% of Fable 5's peak at half the cost; on OSWorld 2.0 it surpasses Fable 5's best at just over a third of the cost; on Zapier's AutomationBench it posts a 100% pass rate. The message to a buyer isn't "this is the frontier." It's "this is good enough for almost everything you do, and it's cheap."
This is price-as-selection, and I've argued the general case before in Your Price Selects Your Customers: a price is a filter, not just a number. Fable 5's $10/$50 selects for the customer who has a job where the frontier edge pays back — and selects out the price-sensitive majority, deliberately. Opus 5's $5/$25 does the opposite: it's set to capture volume and to sit at or below the level where open-weight models and rival labs are undercutting the middle of the market. You don't set these two prices by mistake. You set them because you've decided which customer each model is for.
The floor is falling, so hold the ceiling and drop a strong middle
The macro context makes the segmentation close to forced. I've written about the mechanism in The Inference-Cost Collapse Is About to Break Every AI Pricing Model: the marginal cost of serving a token is falling fast, and open competitors are landing at roughly half of frontier prices. When your own input cost is dropping and someone credible is sitting at half your number, defending a single high price across your whole range is untenable — you either concede the middle or you protect margin at the top while meeting the market below it.
Holding Fable 5 at a premium while dropping a genuinely strong Opus 5 beneath it is the rational response to exactly that squeeze. The high price stays where a small set of buyers will pay it for the hardest work. The cheaper strong model becomes the wall against commoditization in the middle — good enough that a customer has no reason to defect to an open model to save money, because the price gap has already been closed. Opus 5's cache and batch economics reinforce the volume play: cache hits at $0.50/$1, and the Batch API at $2.50/$12.50 — half the standard sticker again for workloads that tolerate latency. That's pricing built to soak up throughput, not to skim.
There's a technical dividend that makes the cheap tier stickier than price alone would. Opus 5 adds self-verification — Anthropic describes it opening pages in a browser at desktop and mobile widths to catch its own layout bugs, iterating until it succeeds. It also carries an alignment score of 2.3 on overall misaligned behavior (the best of recent Claude models), roughly 85% fewer cybersecurity classifier interventions than Fable 5, and support for zero data retention — where Fable 5 mandates 30-day retention. For a regulated or enterprise buyer, the cheaper model is also the lower-friction one to deploy. That lower friction is part of what makes $5/$25 defensible as the default, not an accident of engineering.
The caveat the pricing is quietly admitting
Here's the part that deserves skepticism rather than applause. A vendor pricing a cheaper model above its own benchmark-topping one is telling you something about what it can no longer charge for.
The benchmark that most invites the inversion narrative — Opus 5 at 43.3% versus Fable 5's 33.7% on Frontier-Bench v0.1, said to more than double Opus 4.8 and surpass all other models — is Anthropic's own composite. So are the GDPval-AA v2 SOTA claim (1,861 points) and the Life Sciences gains (organic chemistry +10.2 points over Opus 4.8, protein function +7.7). These are vendor numbers, marketing until independently reproduced. When a lab introduces a fresh internal benchmark ("Frontier-Bench v0.1") and its cheaper new model tops the pricier one on it, the correct response is to scrutinize the benchmark's construction, not to conclude the cheap model won. The third-party figures (CursorBench, OSWorld) are more persuasive precisely because Anthropic didn't build them.
And Anthropic is candid that Fable 5 still leads on the hardest tasks and long-horizon work, and that Opus 5's biology safeguards — looser than Fable 5's but still restrictive — limit long-running autonomous research. So the inversion is not "the cheap one is better." It's "the cheap one is good enough for the middle, and the frontier edge has been narrowed to a small set of jobs." That's the real tell: if raw capability were still the scarce, defensible thing, Anthropic would price the more capable model higher and let capability sort the market. Instead it charges a premium for a use case — unsupervised, days-long autonomy — and prices the capability itself close to commodity. The pricing is admitting that capability is abundant and reliability-at-a-price-point is what's now for sale.
Four models in two months — Opus 4.8, the June Claude 5 wave, then Opus 5 as the fourth — is a cadence that also complicates any buyer's planning. Prices and defaults that move this fast are hard to build a cost model around, and worth naming as a risk, not just a marvel.
What to actually do with this
Model your AI unit economics on the Opus 5 tier — $5/$25 — as your baseline, not the flagship. Anthropic has told you where the default is by making Opus 5 the default on Max and the strongest on Pro; price your workload accordingly. Then treat Fable 5's $10/$50 as an exception budget: a line item reserved for the specific hard, long-horizon, unsupervised tasks where it demonstrably leads, gated by an actual decision, not a habit of reaching for the most expensive model. Price your whole workload at the flagship rate and you overstate your costs by about 2x on sticker before token efficiency even enters — and you'll make build-vs-buy calls on numbers that were never meant to describe your average request.
The counterintuitive move was never really Anthropic making the newer model cheaper. It was Anthropic telling you, through the price tag, that the expensive thing to buy is no longer intelligence — it's the narrow slice of intelligence that runs unattended for days without breaking.