Anthropic just launched Claude Opus 5 without raising the price a single cent over Opus 4.8, and still positions it half a step from the most expensive model in its catalog. Here's what actually changed.
Claude Opus 5 launched on July 24, 2026 keeping the exact price of Opus 4.8: $5 per million input tokens and $25 per million output tokens. What changes isn't the rate but the model's position within Anthropic's catalog — Opus 5 reaches an intelligence level close to Claude Fable 5 (the priciest flagship model) at half its cost, according to figures published by Anthropic and confirmed by independent coverage from TechCrunch and SiliconANGLE.
For teams already weighing the Sonnet-vs-Opus tradeoff in production, this shifts the break-even point: Opus 5 becomes viable for cases where Opus's cost previously ruled it out entirely.
The architecture change most relevant to developers is the effort toggle (low/medium/high) that lets you decide, per call, how much compute to dedicate to reasoning before responding — the same `effort` mechanism we already covered for Opus 4.5, now refined with better separation between the three levels.
At `low`, Opus 5 keeps competitive quality on short tasks at a fraction of the reasoning-token spend — the same effort-tiering logic we recommended in our article on cutting LLM costs, now with a model that makes that tuning more granular.
According to figures published in Anthropic's system card (July 24, 2026) and reported by TechCrunch, Opus 5 more than doubles Opus 4.8's score on Frontier-Bench v0.1, outperforming every other evaluated model. On ARC-AGI 3 — a benchmark specifically designed to resist memorization — Opus 5 triples the score of the next-best model. On OSWorld 2.0 (computer control, real software-usage tasks), Opus 5 beats Fable 5's best result at just over a third of its cost.
These aren't marginal benchmark bumps — they're signals that Anthropic specifically prioritized agentic tool-use and computer-control tasks in this training cycle, an area where it competes directly with the Computer Use capabilities Google just natively integrated into Gemini 3.6 Flash.
Opus 5 keeps the 1-million-token context window and raises the output limit to 128K tokens. Its knowledge cutoff is May 2026 — the most recent of any Claude model released to date, which matters concretely for tasks where the model needs to be current on frameworks, APIs, or recent events without relying exclusively on context injected via RAG.
Anthropic describes Opus 5 as its most aligned model to date, with the lowest rates of deceptive behavior measured internally. A relevant contractual change for companies with strict compliance requirements: Opus 5 supports zero data retention by default, while Fable 5 still requires a 30-day retention window. For customers in regulated sectors (financial, healthcare, legal) evaluating which model in the Anthropic family to use under confidentiality clauses, this policy difference can weigh as much as the benchmarks.
If your team already has an integration with Opus 4.8 via API, Bedrock, or Vertex AI, migrating to Opus 5 is a `model_id` change with no budget adjustment — same price, same billing structure. The practical recommendation is to run the same two-week pilot we suggested for Opus 4.5: measure cost per resolved task, human rework rate, and p95 latency with effort tuned per request type, before moving full production traffic.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel