Executive Summary
On 24 July 2026 Anthropic released Claude Opus 5 at $5 per million input tokens and $25 per million output — the same price as Opus 4.8, half the price of Fable 5, and roughly level with OpenAI's GPT-5.6 Soul. Anthropic's own framing was careful: a model that "comes close to the frontier intelligence of Fable 5 at half the price," positioned as the default on Claude Max and the strongest model on Claude Pro (Anthropic).
The published benchmark table was less careful than the framing. Opus 5 scored 43 on Frontier Bench against Fable 5's 33, took ARC-AGI-3 from a prior best of roughly 8% to 30%, gained about 100 points on GDPval, and edged BrowseComp from 87% to 90% — while conceding ground on a legal benchmark (13.3 to 11.7), on HealthBench Professional, and marginally on DeepSui. On the aggregate, a company's cheap tier beat its expensive tier. That is not how a price ladder is supposed to behave.
The more consequential thing Anthropic did was not the model. It was the chart. The launch materials led with cost-per-task Pareto curves — score against dollars per task, not score against dollars per million tokens. A vendor that prices by the token choosing not to compete on the token is conceding that per-token price has stopped being a meaningful unit of comparison. Every independent evaluation that landed in the following ten days either confirmed that framing or exploited it, and the ones that exploited it found real limits: LlamaIndex measured Opus 5 at roughly 8 cents per page on document parsing against a specialist at a sixth of that, and found that Anthropic's highest thinking mode underperforms its second-highest on a fifth to a third of Anthropic's own reported benchmarks.
The result is a market where the correct question is no longer which model is best, and not even which model is cheapest, but which model finishes your specific task for the fewest dollars — a number no vendor can publish on your behalf, because it depends on your workload. That shift is the subject of this report.
The Metric Changed Before the Model Did
The clearest articulation of the cost-per-task argument in this cycle did not come from a lab. It came from Matthew Berman, whose case rests on Artificial Analysis data showing Kimi K3 at $0.95 per completed task and GPT-5.6 Soul at $1.04 — near-identical, despite Kimi's per-token price being roughly half (Matthew Berman). Kimi consumes about twice the tokens to reach the same result, so the discount is fully absorbed before it reaches the invoice. Across a full benchmark suite the totals were approximately $2,700 for Kimi K3, $2,800 for GPT-5.6 Soul, and $5,600 for Fable 5.
Berman's term for the underlying property is intelligence density — how much reasoning is packed into each token. It is a useful frame precisely because it explains why the two axes came apart. Per-token price measures what you are charged for a unit of output. Cost per task measures how many units the model needs. A model can win decisively on the first and lose on the second, and buyers have been reading only the first.
Anthropic's engineering statements track the same axis. Alex Albert described the headline goal as token efficiency rather than raw capability: the team "put a ton of work into making this model token efficient across domains while still raising the intelligence bar," and he reports preferring Opus 5 over Fable 5 for many coding tasks (Alex Albert). Prompt Engineering's hands-on review found the effort-versus-performance plots showing better price for equal or higher performance than Opus 4.8, with costs climbing only at the highest effort levels (Prompt Engineering).
Berman's operational recommendation follows from the arithmetic rather than from any model preference. He proposes splitting a coding workflow into three stages and assigning a different model to each: plan with the strongest available model, which reads the whole codebase and is input-heavy; execute with a cheap, fast, competent coder, which is output-heavy; review with a different frontier model, input-heavy again. His worked example puts Fable-only at $81, GPT-5.6-only at $46.50, and the mix at $25.55 for equal or better output. The structural insight underneath is that output tokens cost roughly six times input tokens, and the output-dominated stage — writing code from an existing plan — is the easiest stage. It is therefore exactly the stage that should not be running on a frontier model.
Two caveats sit inside the same argument and are easy to lose. First, latency is a cost line: a model that is half the price but takes twice the wall-clock time is not cheaper in any workflow with a human waiting on it. Second, the cross-model review step is not merely a cost dodge — sponsor data from Greptile, disclosed as such, indicates that a Claude-authored pull request reviewed by GPT-5.5 surfaced more bugs than Claude reviewing its own work, with error classes clustering by model. That is an argument for heterogeneity on quality grounds that happens to also be cheaper.
The Players and the Broken Ladder
Three models now sit on the price-performance Pareto frontier: Opus 5, GPT-5.6 Soul, and Kimi K3. The third is open-weights — a 2.8-trillion-parameter Moonshot model at $3 per million input and $15 output. An open model holding a frontier position on cost efficiency, rather than on capability alone, is the genuinely new fact in this arrangement, and it changes the negotiating position of every buyer downstream.
The ladder inside Anthropic's own catalogue is where the inversion bites hardest. Fable 5 is not being retired; it becomes a permanent part of Max and Team Premium subscriptions, but capped at 50% of normal weekly usage limits and excluded from Pro and Team Standard, which retain credit-based access plus a one-time $100 credit (Anthropic, Lev Selector). Anthropic acknowledged that demand for Fable "has been challenging to predict" and that the staged rollout "has been frustrating." Read alongside a benchmark table where the cheaper model wins, the rationing is coherent rather than contradictory: Fable 5 is being metered because it is expensive to serve, not because it is better at the median task.
Anthropic's own positioning language reflects the awkwardness. Its recommendation is to treat Opus 5 as the daily driver and pair it with Fable for planning, brainstorming, or the hardest bugs — a hedge that sits uneasily against the numbers it published on the same day. Prompt Engineering drew the structural conclusion: if an Opus-class model can surpass a Fable-class model in a single iteration, then Fable- and Mythos-class models may stop being consumer products at all and become a restricted category, with the consumer-facing naming staying simple.
Third-party coding benchmarks broadly support the inversion while narrowing it. Cursor's evaluation puts Opus 5 on max close to Fable 5 on max at less than half the price, and on high it surpasses Fable 5 outright; Cognition's Frontier Code places the two on par. Opus 5 beats GPT-5.6 Soul on max while itself running only at extra-high. Those are meaningful margins, but they are margins measured on coding and agentic work, which is where all three vendors have concentrated. Anthropic and OpenAI are both converging away from multimodality toward coding and analytical work, leaving the multimodal lane to Google and the Chinese labs by default.
Two structural notes belong here. Anthropic shipped automatic API fallbacks alongside the model: requests flagged by safety classifiers can auto-route to a different model, and the caller pays the fallback model's price. Users were reporting that a cyber-adjacent refusal silently produced Opus 4.8 within hours of launch. Separately, classifier intervention is expected to be roughly 85% lower than on Fable 5, which is a quiet admission that Fable 5's over-refusal was a known product problem, not a rumour.
What the Independent Benchmarks Found
The independent evaluations that landed within days of launch are the most useful material in this cluster, because they disagree with each other in instructive ways.
The Vals AI index — a GDP-weighted aggregate of CorpFin, Finance Agent V2, SWE-Bench Verified, Terminal-Bench 2.1 and VibeCodeBench — debuted Opus 5 at number two, with slightly lower accuracy than the leader at roughly 20% lower cost. Box ran document-grounded tasks across twelve industries and measured Opus 4.8 to Opus 5 moving from 63 to 78 on the full dataset, 65 to 76 on due diligence, and 67 to 69 on report drafting, with expert review flat and data analysis up six points; the gains concentrate in exhaustive multi-step analysis. Box's evaluation was a disclosed paid placement in Berman's coverage, and it ran no comparison against GPT-5.6 Soul — a gap Berman flagged himself. Greg Kamradt of the ARC Prize called it the most impressive model his team has seen, noting that the ARC-AGI-3 leaderboard's y-axis, capped at 20%, would have to be redrawn.
The counterweight is LlamaIndex's ParseBench pass, and it is specific rather than sour. On document understanding Opus 5 is roughly on par with Opus 4.8 — a few points worse on dense tables, slightly better on charts and visual grounding — at about 8 cents per page. Gemini 3.6 Flash beats it on tables at half that price; LlamaParse in agentic mode beats it on all fronts including tables at a sixth (LlamaIndex). The demonstration was pointed: running Anthropic's own 193-page Opus 5 system card, dense with labeled and unlabeled charts, through both, LlamaParse parsed unlabeled line and bar charts at near-perfect accuracy for 1.25 cents per page, while Opus 5 parsing its own PDF showed a 20-30% reduction in chart accuracy and cost between 8 and 33-plus cents per page depending on thinking mode. The recommendation was explicit and worth quoting for its shape: use Opus 5 as much as you want for coding and knowledge work, but not for document parsing at scale.
That is the cost-per-task argument turned back on the vendor that popularised it. A frontier generalist and a task-specific parser are not competing on intelligence; they are competing on dollars per finished page, and on that axis the generalist loses by a factor of six.
The second finding from the same source is more troubling for anyone building a routing layer. Across roughly 20-30% of the benchmarks Anthropic reported, max thinking underperforms xhigh (LlamaIndex). This is not a rounding artifact — Berman independently observed that on Frontier Bench, Opus 5's score dropped at its highest thinking level while costing more. Whether that is an emergent property of stronger models or a post-training issue is genuinely open. The operational consequence is not: the highest setting is not the safe default, and any system that escalates thinking level on failure is escalating into a regime where it may pay more for a worse answer.
One credibility note deserves recording without being inflated. Prompt Engineering caught an error in Anthropic's own launch materials, where on agentic coding a score of 53.4 was presented as higher than 53.5. The delta is immaterial; the reviewer's framing — that this was not expected from Anthropic — is the part worth keeping, because launch-day charts are now the input to procurement decisions made within hours.
The Mechanism: A Flagship That Became a Training Asset
The most credible explanation for how a cheaper model beat a more expensive one is that the expensive one made it. The hypothesis circulating across several independent channels is that Opus 5 is a distillation of, or was trained by, Fable 5 — a second-generation optimisation pass over a frontier training run. Nathan Lambert's public read supplies the mechanism, describing insane numbers for Opus 5 driven by faster iteration speed and scaled reinforcement learning, with Fable "too big to RL as well yet."
If that holds, the industry pattern to expect is a change in what a frontier run is for. Train an oversized model, then harvest it into a cheaper deployable one. The expensive model becomes a training asset rather than a product — which is precisely the posture Anthropic's rationing of Fable 5 already implies, and it explains why the flagship can be capped at 50% of weekly limits without the vendor appearing to abandon its own top tier.
A second-order result is the safety and capability decoupling, which is the anomalous finding of the launch. Opus 5 is stronger than Opus 4.8 on cybersecurity tasks overall, yet scores 4 on exploitation success where Mythos 5 scores 13 and Opus 4.8 scored 0. Capability was selectively removed without the across-the-board degradation that usually accompanies guardrails. Prompt Engineering credits that deliberate reduction with letting the model ship on schedule rather than waiting on government review — a consumer benefit produced by a safety decision, which is not the usual direction of that trade.
The capability itself is best evidenced not by benchmarks but by what practitioners built in the first week. Andrej Karpathy handed the model the first paragraph of The Lord of the Rings, a 1M-token budget costing roughly $10, and asked for a three.js render; it ran about two hours and produced 5,500 lines of code that procedurally render the story (Andrej Karpathy). His economic reading is the relevant one: nobody in their right mind would hand-write something that custom, so model stamina moves the work from no one would ever do this to sure, why not, it is roughly free. He also names the limit the domain exposes — the model cannot efficiently perceive video or play, so it audited its own work by taking screenshots painstakingly and still produced jank.
At larger scale, a build using Matt Shumer's Gauntlet Loop method produced Modern Claudefare, a browser-native Call of Duty clone at 84,100 lines with four maps and voice-chat multiplayer, in a few days on Opus 5 in High mode (Matt Shumer). The method matters more than the artifact: give the agent a bar it cannot talk its way around, split the work across sub-agents, and never let the builder grade itself. That is a verification loop bolted onto a cheap model — and it is the same trade the cost-per-task argument keeps producing from different directions. You buy correctness with structure rather than with model tier.
What This Means for Anyone Buying Tokens
The commercially sharpest treatment of this in the cycle came from Sarah Sachs, who negotiates Notion's model contracts, and her framing generalises beyond Notion (AI Engineer). Her central fact is that your supplier is your competitor: you buy tokens at a markup from an entity serving the same capability as a first-party product at a very different cost of goods, then resell them at another markup. And she names two pricing events Notion hits roughly monthly, both of which silently break unit economics — a reasoning model upgraded at identical per-token pricing that consumes three times the output tokens, and a version bump at 40% more than its predecessor which is then deprecated in four months.
Both are cost-per-task events invisible on a per-token price sheet. Her operative instruction is to reason about cost per capability per second rather than about tokens, to build for multi-model from the start because it is hard to kill the cache and switch models mid-transcript, and to treat price as a poor quality signal — if you are slightly behind the frontier you only need to be about $1 per million cheaper to take the rest of the market, so price reflects competitive position, not capability.
Her cheapest optimisation is the one most easily forgotten in a report about model selection: do not use a model. Notion ships deterministic workers for exactly this, on the reasoning that you do not need an LLM to turn a CSV into a PDF, to call a tool that has a CLI, or to run a deterministic SQL query. Misrouting is expensive in both directions — large-scale data analysis warrants an Opus-class model, and triaging an email inbox on one is, in her words, ripping off both the customer and yourself.
The subscription tier tells the same story from the practitioner's side. The operator behind the BridgeMind stream runs two Claude Max subscriptions, sat at 84% of his weekly Fable 5 limit, and downgraded his Codex tier — and his stated reason for wanting Opus 5 was headroom, not intelligence (BridgeMind). That is exactly what Anthropic priced the model for, and it identifies the binding constraint on agentic work today as usage limits rather than capability.
The same source supplies the most useful contradiction in this cluster. Benchmark-side, Kimi K3 sits on the cost-per-task frontier; practitioner-side, running production workloads at 18-25 tokens per second with 8.7 seconds of latency, his verdict is not to use it as a main driver because it is too slow. Both are true, and the reconciliation is that a cost-per-task chart cannot see a serving bottleneck. Yet the same model's swarm mode was the most-praised capability in his stream: 20-30 sub-agents at once, 70-plus concurrent agents at only 33% of a five-hour limit, against his estimate that 30 Fable 5 sub-agents would not get through a single prompt. Slow-but-abundant and fast-but-rationed are different capability classes, not substitutes, and parallel breadth can beat single-stream speed on wall-clock time to a finished task.
Finally, a bias correction worth building into any evaluation process. Factory AI's Eno Reyes put it plainly: if he does not know which model he is using, he says it is great; if he knows it is the frontier model, his bias kicks in. Blind evaluation is not methodological fussiness here — it is the only defence against paying a premium for a label.
Outlook
The near-term direction is legible. Cost per task is now the axis vendors compete on, and because it is workload-dependent, the number that matters cannot be published by a vendor at all. The practical consequence is that buyers who do not run their own evaluation are, structurally, taking the vendor's workload mix as a proxy for their own — and the ParseBench result shows how far apart those can be, with a factor-of-six gap on a task the frontier model handles competently but not economically.
Three things are worth watching. Whether the distillation pattern holds — that is, whether the next flagship from any lab ships primarily as a training asset with a cheaper harvested model as the consumer product — will determine if the tier inversion was a one-off or the new shape of a release cycle. Whether the non-monotonic thinking-level result reproduces across vendors will determine how routing layers should escalate on failure; today the honest answer is that escalating to maximum effort is not reliably an improvement and is reliably more expensive. And whether open-weights models hold their Pareto-frontier position matters beyond price: as Notion's account makes clear, a credible open alternative is negotiating leverage against an oligopoly of two or three, and that leverage does not require anyone to actually deploy it.
What has already changed is smaller and more durable. A vendor that prices by the token led its own launch with a chart that prices by the task. That concession is not reversible, and it hands buyers a question they now have to answer for themselves.
