Claude Opus 5 Promises Half the Price of Fable 5 — Real-World Testing Found Only 20% in Savings

On July 23, 2026, Anthropic released Claude Opus 5, positioning it as a model that delivers near-Claude Fable 5 intelligence at exactly half Fable 5's price — $5 per million input tokens and $25 per million output tokens, the same rate as the outgoing Opus 4.8. The model is live across Claude.ai, the Claude API, Claude Code, AWS Bedrock, and Google Vertex AI, and became the new default on Claude Max the same day. Independent benchmarks back up the intelligence claim. What they don't settle is the cost claim: within 24 hours, a widely-followed developer testing the model in production found that real-world savings land closer to 20%, not 50% — because Opus 5 tends to use more tokens to finish the same task.

A Frontier Model at Half the Sticker Price

Opus 5 ships with a 1-million-token context window, a 128K-token max output, and a May 2026 knowledge cutoff — the most current of any model in Anthropic's lineup. Standard pricing holds at $5/$25 per million input/output tokens, unchanged from Opus 4.8; a faster mode that runs roughly 2.5x quicker doubles that to $10/$50. By comparison, Fable 5 — Anthropic's most capable and most expensive model — costs meaningfully more per token, which is the entire basis for Anthropic's 'half the price' framing.

Where Opus 5 Actually Wins on Benchmarks

The independent benchmarks are consistent across outlets that tested it: on Frontier-Bench v0.1, an agentic coding evaluation, Opus 5 scored 43.3% against Fable 5's 33.7% — more than double Opus 4.8's prior score. On CursorBench 3.2, Opus 5 landed within 0.5% of Fable 5's peak result at half the per-task cost. On ARC-AGI 3, a reasoning benchmark, it scored roughly three times higher than the next-best model, and on OSWorld 2.0, a computer-use benchmark, it beat Fable 5's best score while spending just over a third as much. On the independent Artificial Analysis Intelligence Index, Opus 5 topped the field at 61, and led the Agentic Index at 55.3 — both ahead of Fable 5 and GPT-5.6 Sol.

The Catch: Theo's Real-World Test Found Only 20% Savings

Those numbers are why Anthropic can credibly call Opus 5 a frontier model at half the price. But developer Theo, founder of t3.gg and one of the most closely watched voices in the coding-agent community, spent a full day running real production tasks through it and reported a narrower result: actual savings closer to 20%, not the 50% the sticker price implies. His explanation is straightforward — Opus 5 is cheaper per token, but it tends to use more tokens per task than Fable 5 does to reach the same outcome, whether through longer reasoning traces, more tool calls, or more verbose intermediate steps. Half the per-token price times more tokens per task nets out to a much smaller bill reduction than the headline number suggests.

Why Cheaper Per Token Doesn't Mean Cheaper Per Task

This is a distinction that per-token pricing tables consistently obscure: the metric that determines your actual invoice is cost per completed task, not cost per million tokens. A model that's 50% cheaper per token but burns 60% more tokens on average for the same class of task can still leave you with a smaller but real bill increase, not the discount the pricing page implies. It's the same dynamic that showed up when GPT-5.6 Sol's token appetite drew criticism for running up costs on agentic workloads despite competitive per-token rates — cheap tokens and cheap tasks are not the same claim, and vendors have every incentive to advertise the first while teams need to budget for the second.

What It Means for Developers and AI Teams

For teams evaluating which model to route agentic coding workloads to, the practical move is the same one Theo used: don't compare price sheets, run your own representative tasks through both models and compare the actual bill. Opus 5's higher benchmark scores and lower per-token price make it a reasonable new default to test against, especially for teams currently on Fable 5 who feel their token spend is climbing faster than their output — but 'half the price' should be treated as a ceiling on possible savings, not a guarantee. Given Opus 5 is now available through Claude Code, Bedrock, and Vertex AI as well as the direct API, most teams can run that comparison without any migration cost beyond a config change.

Bottom Line

Anthropic's pricing claim for Opus 5 is technically accurate and independently verified on a per-token, per-benchmark basis — it is a frontier-class model at half of Fable 5's rate card. What the first wave of real-world testing adds is the reminder that per-token price and per-task cost are different numbers, and the gap between them is exactly where AI infrastructure budgets tend to go wrong. Teams that measure actual task cost before switching will get a real, if more modest, discount; teams that switch on the headline number alone may find their bill dropped by 20%, not 50% — and either way, on production agentic workloads, that gap is worth measuring rather than assuming.