Claude Opus 5 and the Effort Dial: 2026’s Model Wave Trades Cost for Capability

Anthropic’s Fourth Release in Under Two Months

On July 24, 2026, Anthropic introduced Claude Opus 5, a model the company is positioning as its “everyday” workhorse for enterprises, knowledge workers, and developers rather than a pure frontier flex. Pricing did not move: Opus 5 bills at $5 per million input tokens and $25 per million output tokens, identical to its predecessor Opus 4.8. What changed is how much computation a team can choose to spend answering any given prompt. According to Axios’s coverage of the launch, Opus 5 is Anthropic’s fourth Claude 5-family model shipped in under two months — after Mythos 5 and Fable 5 in June, followed by Sonnet 5 — a cadence Axios frames as evidence that AI deployment has shifted from blockbuster launches toward rapid, continuous tuning of capability, cost, and speed.

The headline claim, repeated across Anthropic’s own materials and independent coverage, is that Opus 5 “comes close” to the intelligence of Anthropic’s top-tier Fable 5 model on many tasks while costing half as much. Fortune’s reporting on the release notes the launch directly answers “growing concerns from enterprise customers about expensive AI bills,” and that Anthropic still recommends Fable 5 for the hardest, most autonomous multi-day projects. Opus 5 is instead being pitched as the model for the other 90% of enterprise work — support automation, internal tooling, document analysis — where near-frontier quality at a controllable price matters more than squeezing out the last few points of benchmark performance.

How the Effort Dial Actually Works

The mechanism drawing the most attention is Opus 5’s effort parameter. Coverage of the release describes five effort levels — low, medium, high, xhigh, and max — with high set as the default when a developer omits the parameter. Crucially, the per-token price does not change across effort levels; what changes is how many thinking and tool-call tokens the model consumes before it answers. A short, well-specified prompt at low effort might resolve in a handful of tokens, while the same prompt at max effort could trigger extensive internal reasoning before Claude commits to a response.

One analysis of the effort parameter reports that, at a volume of 100,000 monthly requests, shifting a workload from high to medium effort produced roughly a 44% cost reduction, and dropping from high to low saved closer to 66% — though the site is careful to note these figures are workload-dependent, not guaranteed. There is also a hard technical boundary: thinking cannot be disabled once a request is set to xhigh or max effort, and Anthropic’s API returns a 400 error if a caller tries. Anthropic’s own guidance, per the same coverage, is to start new integrations at high effort, then step down to medium or low once evaluation results show quality holds — or step up to xhigh or max only for genuinely demanding work. On top of the effort dial, Opus 5 also ships a Fast mode that runs at roughly 2.5 times the default speed for twice the base price, giving teams a second lever — speed versus cost — independent from the reasoning-depth lever.

Benchmark gains reported by Anthropic reinforce the “near-frontier for less” pitch: Opus 5 more than doubles Opus 4.8’s score on Anthropic’s internal Frontier-Bench v0.1 at a lower cost per task, lands within 0.5% of Fable 5’s peak score on CursorBench 3.2 at half the cost, and surpasses Fable 5’s best recorded result on the OSWorld 2.0 computer-use benchmark at just over a third of the cost. Developer Simon Willison’s independent write-up highlights a specific example from Anthropic’s materials: given an image it could not directly analyze, Opus 5 wrote its own computer-vision pipeline to extract 3D geometry and reconstruct a FreeCAD model — the kind of improvised, multi-step problem solving Willison frames as evidence of the model “becoming more generally capable,” not just narrowly trained. Willison also notes a real limitation: Opus 5 approaches Anthropic’s Mythos 5 at finding cybersecurity vulnerabilities but “remains substantially behind” it at exploiting them, a gap Anthropic left in place deliberately by not training Opus 5 on offensive cyber tasks.

Three Labs, Three Different Knobs

Opus 5 did not land in isolation. It arrived roughly two weeks after OpenAI’s GPT-5.6 reached general availability on July 9, 2026, and in the same month Google pushed out a new generation of cost-tuned Gemini models. All three labs are converging on the same underlying bet — that 2026’s competitive edge is not just raw intelligence but the ability to trade intelligence for cost on demand — but each is solving it differently.

OpenAI split GPT-5.6 into three separate priced tiers rather than one dial: Sol at $5 input / $30 output per million tokens, Terra at $2.50/$15, and Luna at $1/$6, according to Artificial Analysis’s benchmark breakdown. On Artificial Analysis’s Intelligence Index, Sol at maximum effort scores 59 points — one point below Fable 5’s 60 — while costing roughly $1.04 per completed task, described in the same analysis as “approximately one third of the cost” of matching Fable 5’s intelligence. On the Coding Agent Index, Sol at max scored 80 points, leading all three tiers evaluated, at around 40% less cost than Fable 5. Where Anthropic asks developers to turn one dial, OpenAI is effectively asking them to pick a different model per job.

Google took a third approach. Its July 2026 update introduced Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, aimed squarely at production AI agents rather than frontier benchmarks. Gemini 3.6 Flash prices at $1.50 per million input tokens and $7.50 per million output tokens, a 16.7% cut to output pricing versus Gemini 3.5 Flash, and ships with four thinking levels — minimal, low, medium, and high — that let engineering teams route simple, high-volume subagent calls to cheap minimal-effort inference while reserving high-effort reasoning for genuinely multi-step tasks. The same coverage reports the model uses roughly 17% fewer output tokens than its predecessor on comparable tasks and that effective cost per completed task falls by up to 71% on agentic coding workloads specifically.

What Enterprise AI Buyers Should Do About the Effort Dial

1. Rebuild vendor evaluation around per-task cost curves, not sticker price

A flat $5/$25 rate card tells you almost nothing about what Opus 5 will actually cost your workload, because the effort setting — not the token price — is now the main cost lever. Before signing or renewing a contract, run your actual production prompts at each effort level (low through max) and plot cost against output quality for your specific use case. Don’t rely on vendor-published benchmark averages; a customer-support triage workload and a legal-document review workload will land on completely different points of the same curve.

2. Assign a default effort level per workflow, not per organization

Treat effort level as an engineering decision made at the workflow layer, the same way you’d choose a database index. High-volume, low-stakes tasks (ticket routing, summarization of short documents) belong at low or medium effort; anything touching financial figures, contracts, or customer-facing accuracy claims should default to high or xhigh. Anthropic’s own guidance — start at high, step down only where evaluations show quality holds — is a reasonable default policy to adopt wholesale rather than reinvent internally.

3. Build a live cost-per-task dashboard before scaling any agentic workflow

Because thinking and tool-call tokens (not just visible output) drive spend at higher effort levels, a single misconfigured agent looping at max effort can burn budget invisibly for days before anyone notices. Instrument cost-per-completed-task as a first-class metric alongside latency and error rate, and set hard budget alerts per effort tier — not just per API key — before any agentic deployment goes to production scale.

4. Treat the three-lab divergence as a hedge, not a decision to make once

Anthropic’s single-model dial, OpenAI’s three-tier Sol/Terra/Luna split, and Google’s four-level Gemini 3.6 Flash thinking budget are three different bets on how developers want to manage the cost-capability tradeoff. Don’t lock a multi-year architecture to one lab’s approach. Build an abstraction layer that can route a given task to whichever provider’s cost curve fits best this quarter — the three approaches are different enough that the cheapest option for a given task type will keep shifting as each lab iterates.

Where This Fits in 2026’s Model Economy

The effort dial is less a feature than an admission: after several years of labs racing purely on intelligence, 2026’s frontier competition is now openly about the price of intelligence per task. Opus 5, GPT-5.6, and Gemini 3.6 Flash all launched within roughly three weeks of each other pursuing the same insight from three different angles — one dial, three price tiers, four thinking budgets — because all three labs have concluded that most enterprise AI spend is not going toward the hardest 1% of tasks that need frontier reasoning, but toward the routine 99% that needs “good enough, cheaply, at scale.” The labs that win the next phase of enterprise adoption will likely be the ones whose cost-tuning mechanism is easiest to operationalize inside existing engineering workflows, not necessarily the ones with the highest peak benchmark score. For buyers, that means the real competitive question shifts from “which model is smartest” to “whose cost dial is cheapest to manage at our scale” — a much less glamorous question, but the one that will actually show up on next year’s cloud bill.

Frequently Asked Questions

Sources & Further Reading

• Introducing Claude Opus 5 — Anthropic
• Introducing Claude Opus 5 — Simon Willison
• Anthropic releases Claude Opus 5: Here’s how it’s different than what’s already out there — Fortune
• Anthropic releases new model, Opus 5 — Axios
• Claude Opus 5 effort levels: cut token spend — ecorpit
• GPT-5.6 benchmarks across Intelligence, Speed and Cost — Artificial Analysis
• Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — Google AI Updates, July 2026
• Google’s Gemini 3.6 Flash targets enterprise agent token costs — AI News