Claude Opus 5 Halves the Price of Frontier Coding

Anthropic just made its most capable coding model roughly half as expensive to run without cutting the sticker price. On July 24 the company introduced Claude Opus 5, a model it says comes close to the frontier intelligence of Claude Fable 5 “at half the price,” priced at $5 per million input tokens and $25 per million output tokens—identical to its predecessor, Opus 4.8. The savings do not come from a discount. They come from the model completing the same work with far fewer tokens and turns, which is where the real cost of agentic coding lives.

That distinction is the whole story for operators. On Anthropic’s own benchmarks, Opus 5 more than doubles Opus 4.8’s score on the Frontier-Bench coding evaluation at a lower cost per task, and on CursorBench 3.2 it lands within 0.5% of Fable 5’s peak score at roughly half the cost per task. Independent-facing endorsements echo it: Cursor reported the model performs “just under Fable 5” with similar behaviors, and Cognition’s Devin team said it approaches Fable-level performance on FrontierCode 1.1 at half the cost, with particular strength on debugging and root-cause analysis. The model is already available in GitHub Copilot for Pro+, Max, Business, and Enterprise tiers, billed at API list price. VentureBeat framed the launch as a shift from raw capability to the economics of daily use, and CNBC noted Anthropic is appeasing a cost-conscious enterprise base less inclined to experiment without clear ROI.

Why cost-per-task is the number that matters

Sticker price per token has always been a misleading unit for agents, and Opus 5 makes the point unmissable. An autonomous coding run reasons, edits, tests, retries, and verifies before a human sees a result, so the invoice is set by total tokens consumed, not the rate card. Here is the takeaway nobody published cleanly: because Opus 5 holds the $5/$25 rate flat while cutting tokens-per-task toward half on matched-quality work, a team already paying for Opus 4.8 or Fable 5 can expect its per-completed-task bill to fall by a similar proportion with no contract change and no quality regression. One early tester, a trading firm, reported reaching the same benchmark result using “roughly a seventh of the reasoning tokens and under half the latency of Opus 4.8.” That is a step change in effective cost, not a rounding-error improvement—a roughly 50% reduction in what it costs to finish a unit of work, achieved without touching the rate card.

The pattern generalizes beyond coding. On ARC-AGI 3, a test of novel problem-solving, Anthropic reports Opus 5 scoring three times the next-best model; on the OSWorld 2.0 computer-use benchmark it claims to beat Fable 5’s best result at roughly a third of the cost. Anthropic’s Opus 5 system card documents the safety envelope, while its prompting guide shows how effort settings let teams dial intelligence up or tokens down per call. The through-line is that Anthropic has learned to convert the same silicon into more finished work, which is the only cost lever that survives the compute-constrained market Alphabet just repriced with $811 billion in commitments.

The behavioral gains compound the economics. Anthropic emphasizes that Opus 5 verifies its own work—one legal-tech customer measured 26% fewer tokens at max reasoning, another saw “a third fewer turns and tool calls and 60% less time” on finance tasks. Fewer turns means fewer failed loops, and failed loops are where agent budgets quietly bleed out. This is the same efficiency-as-cost-of-goods logic driving Grok’s push to make token efficiency the benchmark and the rationing that followed the industry’s token binge. Opus 5 is Anthropic answering that pressure from the quality end rather than the price end.

What could break the thesis, and who should switch

The counterpoint is release fatigue and vendor lock-in. Opus 5 is Anthropic’s fourth model in under two months, following Fable 5, Sonnet 5, and Mythos 5—a cadence that keeps the enterprise-adoption lead Anthropic built over OpenAI intact but strains teams trying to standardize on a stable target. Benchmarks are also Anthropic’s own; the outside numbers come from partners with commercial incentives, so the half-cost claim deserves a controlled A/B on your real repository before you migrate. There are new guardrails, too: Opus 5 ships stricter cyber classifiers that can block security-adjacent requests and fall back to Opus 4.8, a friction point for security engineering teams and a reminder of the leash regulators and labs keep tightening on frontier models.

Who should switch this quarter? Teams running long-horizon agentic coding on Opus 4.8 or Fable 5 have the clearest case: same price, materially fewer tokens, better verification. Run one week of production traffic through Opus 5 at matched effort settings, compare cost-per-merged-PR rather than cost-per-token, and let the delta decide. If your workload leans on binary vulnerability scanning or penetration testing, test the classifier behavior first—the fallback may erase the savings. For everyone else building on rented inference, Opus 5 is the cleanest evidence yet that the next round of AI cost reduction arrives disguised as a quality upgrade.