Verdict: Claude Opus 5 (released July 24, 2026) keeps the same $5 per million input / $25 per million output as Opus 4.8 — but the adjustable effort dial is the real cost story for enterprises. By routing 80% of routine work through lower effort levels and reserving high/xhigh effort for genuinely hard tasks, teams can cut their per-task AI bill by 40-66% on the same model, without switching to a cheaper tier. The new competitive battleground in enterprise AI is not raw capability or list price — it's cost per task, and Opus 5's effort control is the sharpest lever available.
TL;DR:
- Opus 5 launched July 24, 2026 at $5/$25 per million tokens (identical to Opus 4.8), approaching Fable 5's frontier performance at half the cost (Anthropic)
- The five-level effort dial (low, medium, high, xhigh, max) is the primary cost lever — not the per-token price
- High effort is the API default; teams leaving it there are overpaying on routine work
- A simple 80/20 routing strategy (80% low/medium, 20% high) can cut a 1,000-task/day workload's monthly bill by ~40% (codersera)
- Beta features — mid-conversation tool changes and automatic model fallbacks — further reduce wasted spend on blocked or re-routed requests
- The launch follows OpenAI's GPT-5.6 family (July 9) and Moonshot AI's open-weight Kimi K3 (July 27), intensifying the cost-per-task race
What is Claude Opus 5's effort dial and why does it matter for cost?
The effort dial is a per-request parameter with five settings — low, medium, high, xhigh, and max — that controls how much reasoning compute the model spends before answering. The API defaults to high, which behaves identically to omitting the parameter entirely. Set it as output_config.effort on the request; no beta header is required (Anthropic).
The critical mechanism: reasoning ("thinking") tokens bill as output at $25 per million. A high-effort request generates far more reasoning tokens than a low-effort one on the same prompt. The per-token rate never changes — but the number of tokens consumed swings dramatically. Lower effort means the model combines operations into fewer tool calls, skips preamble, and reaches a conclusion faster. Higher effort means more multi-angle reasoning, more tool calls, and longer deliberation.
This is unlike OpenAI's GPT-5.6 family, which splits into three separately-priced models (Sol, Terra, Luna) where you cut cost by choosing a smaller model. Opus 5 keeps one model and cuts cost by spending fewer tokens through the effort parameter. One gives you a lower price floor; the other gives you finer control on a single quality profile (OpenAI). For a full breakdown of the effort dial's mechanics, see our Claude Opus 5 effort dial and pricing guide.
| Effort level | Relative token spend | Good fit | Bad fit |
|---|---|---|---|
low |
Lowest | Renames, formatting, classification, mechanical edits | Anything requiring a design choice |
medium |
Low | Small features with clear tests, doc updates, research summaries | Cross-service changes, complex debugging |
high |
Moderate (default) | Most production agent work, code review | Trivial edits — pure waste |
xhigh |
High | Hard debugging, migration planning, unfamiliar code | Latency-sensitive interactive paths |
max |
Highest | Benchmarks, genuinely stuck problems, one-off hard analysis | Default routing — this is the expensive mistake |
Source: effort-level guidance synthesized from Anthropic platform docs and Hashnode.
How much does Claude Opus 5 actually cost per task?
The headline price is $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8 and exactly half of Fable 5's $10/$50 rate (Anthropic). But list price is the least interesting number. What determines your bill is how many thinking tokens a task consumes, and that is the effort setting you choose.
Here's a worked cost example on a 40,000-token input prompt:
| Low / medium effort | High effort | |
|---|---|---|
| Input tokens | 40K | 40K |
| Output tokens (incl. thinking) | ~6K | ~20K |
| Input cost | $0.20 | $0.20 |
| Output cost | $0.15 | $0.50 |
| Per-task total | $0.35 | $0.70 |
Same prompt, double the cost — decided entirely by one parameter. The takeaway isn't "always use low." Some tasks genuinely need deeper reasoning to get the right answer on the first pass. The takeaway is that effort should be a deliberate per-workload decision, not a global default.
For a team running 1,000 agentic coding tasks per day:
| Strategy | Daily cost | Monthly cost (30 days) |
|---|---|---|
| All high effort | $700 | $21,000 |
| 80% low/medium + 20% high | $420 | $12,600 |
| Savings | $280/day (40%) | $8,400/month |
In an illustrative model at 100,000 requests per month, moving from high to medium effort cut the bill approximately 44%, and high to low cut about 66% (ecorpit). Real savings depend on your prompts, so measure on your own workload first.
How does Opus 5's cost compare to Fable 5 and the competition?
Opus 5's positioning is precise: it approaches Fable 5's frontier intelligence at half the per-token price. On CursorBench 3.2, Opus 5 at max effort lands within 0.5% of Fable 5's peak score. On OSWorld 2.0 (computer use), Opus 5 outperforms Fable 5's best result at just over a third of the cost (Anthropic). For a deeper head-to-head on the performance trade-off, see our Claude Opus 5 vs Fable 5 half-price comparison.
| Model | Input / 1M tokens | Output / 1M tokens | Context window | Best for |
|---|---|---|---|---|
| Claude Opus 5 | $5 | $25 | 1M | Flagship reasoning + coding, cost-controlled |
| Claude Fable 5 | $10 | $50 | 1M | Maximum capability ceiling |
| Claude Sonnet 5 | $2 (intro) / $3 (standard) | $10 / $15 | 1M | Price/performance balance |
| OpenAI GPT-5.6 Sol | $5 | $30 | — | OpenAI flagship |
| OpenAI GPT-5.6 Luna | Lower | Lower | — | OpenAI cost-efficient tier |
Sources: Anthropic, OpenAI, Claude pricing docs. Sonnet 5 introductory pricing runs through August 31, 2026.
Additional cost levers beyond the effort dial:
- Prompt caching: Cache reads cost $0.50 per million tokens for Opus 5 — a 90% discount on cached input. The cache threshold kicks in from just 512 tokens (aipricing.guru).
- Batch API: Halves both rates to $2.50/$12.50 per million for non-time-sensitive workloads.
- Fast mode: Runs at ~2.5x default speed for double the price ($10/$50 per million). Available on the first-party Claude API only — not through the Batch API.
What are the two beta features that reduce enterprise waste?
Anthropic shipped two beta features alongside Opus 5 that directly address enterprise cost concerns (Anthropic):
1. Mid-conversation tool changes. Developers can now change which tools Claude can use within a conversation — without invalidating the prompt cache. Previously, swapping tools meant re-sending the full conversation context, which re-billed all those input tokens. Now the cached context persists across tool reconfigurations, which means less wasted input spend on long agentic sessions.
2. Automatic fallbacks. Requests flagged by safety classifiers can now reroute to another available model instead of being blocked outright. This matters for cost in two ways: blocked requests that get retried from scratch waste tokens, and aborted work-flows waste the tokens already spent. Auto-fallback keeps the workflow moving — and the billing productive. Anthropic calls Opus 5 "the least susceptible to being tricked into misuse" of any Claude model, and biology-related requests blocked on Fable 5 now route to Opus 5 as a fallback.
How to build an effort-routing strategy for your team
The single most common cost failure: teams set the highest effort globally because it produced the best demo result, then discover a month later that 80% of their spend went to tasks a low-effort call would have handled identically. Here's a practical routing approach:
- Start at high effort. Anthropic recommends beginning at the default
highand validating quality on your representative workload. - Step down on routine work. Move classification, formatting, documentation, simple refactors, and mechanical edits to
mediumorlow. Run an A/B eval: if quality holds, keep the lower setting. - Reserve xhigh/max for genuinely hard problems. Hard debugging, architecture decisions on unfamiliar code, multi-file reasoning, and one-off analytical deep-dives. These are the 20% where the extra reasoning tokens earn their cost.
- Build a lightweight router. A cheap classifier or a few heuristics (prompt length, task type keyword, request source) can bucket each request before it hits the model. This is where the real savings compound.
- Measure cost per task, not cost per token. Track total tokens consumed per completed task at each effort level. A task that succeeds on the first pass at medium effort is cheaper than one that needs two retries at high effort.
Latency is also part of the price equation. At max effort, Opus 5 measures roughly 62.7 seconds to first token — that's thinking time, and it's real. Match the effort to the latency tolerance of the surface:
| Surface | Latency tolerance | Recommendation |
|---|---|---|
| Autocomplete | Under 300ms | Never Opus 5 — use a small fast model |
| Interactive chat | A few seconds | Low/medium, or Fast mode, or Sonnet 5 |
| PR review bot | A minute or two | High effort is fine |
| Background ticket agent | Minutes | xhigh is fine, max for stuck tasks |
| Nightly batch analysis | Hours | Anything goes |
What does "cost per task" mean for the AI competitive landscape?
The Opus 5 launch lands in a crowded few weeks. OpenAI shipped the GPT-5.6 family (Sol, Terra, Luna) on July 9, 2026 (TechCrunch). Moonshot AI released the open-weight Kimi K3 on July 27, 2026 — a 2.8-trillion-parameter model now free to download (Startup Fortune). Both Anthropic and OpenAI are reportedly preparing IPOs later in 2026, adding pressure to ship competitive releases.
The strategic shift this launch signals: AI pricing is increasingly competing on cost per task rather than raw capability. With frontier models now clustered near the top on benchmarks, the differentiator is not "who is smartest" — it's "who delivers the right answer at the lowest total spend." Opus 5's effort dial is Anthropic's answer: one model that scales from cheap-and-fast to deep-and-thorough, controlled at the request level. OpenAI's answer is three models at three price points. The market will decide which model enterprises prefer, but the direction is clear — the cost-per-task race is now the main event.
What this means for you
If you're an engineering team or enterprise already using Claude, the action is immediate: audit your effort settings. If you're leaving everything at the default high, you are almost certainly overpaying on routine work. Start with an A/B eval on your most common task type — step effort down one level and compare quality. For practical agentic patterns that pair well with cost-controlled routing, see our practical agentic playbook for Claude Opus 5. For teams evaluating Opus 5 against GPT-5.6 or open-weight alternatives like Kimi K3, the comparison framework is no longer "which model is best on benchmarks" but "which model gives me the lowest cost per correctly-completed task at my workload mix." The broader shift is part of the AI infrastructure supercycle reshaping how builders ship in 2026. Build the router, track cost per task, and let the data decide.
FAQ
Q: Is Claude Opus 5 cheaper than Opus 4.8?
A: No. The per-token rates are identical: $5 per million input tokens and $25 per million output tokens. The "half the price" framing refers to Opus 5 vs. Fable 5 ($10/$50), not vs. Opus 4.8. The real savings come from the effort dial, which can cut per-task cost by 40-66%.
Q: What is the default effort level on the Claude Opus 5 API?
A: The API defaults to high effort, which behaves identically to omitting the parameter. Anthropic recommends starting at high, then stepping down to medium or low where your evaluations show quality holds, or up to xhigh/max for demanding work.
Q: How do I cut my Claude Opus 5 bill without losing quality?
A: Route the bulk of your traffic through low or medium effort and escalate only the genuinely hard 20% to high or xhigh. A simple classifier or heuristic router can bucket requests by task type before they reach the model. Combined with prompt caching (90% discount on cached input) and the Batch API (50% off for non-time-sensitive work), the savings compound.
Q: What is Claude Opus 5 Fast mode and is it worth it?
A: Fast mode runs at approximately 2.5x default speed for double the per-token price ($10/$50 per million). It increases output throughput rather than reducing time to first token, is available on the first-party Claude API only, and cannot be combined with the Batch API. It's worth it for latency-sensitive interactive workloads where speed matters more than per-token cost.
Q: How does Opus 5's effort dial compare to OpenAI's GPT-5.6 tiers?
A: GPT-5.6 splits into three separately priced models (Sol, Terra, Luna) — you cut cost by choosing a smaller model. Opus 5 keeps one model and cuts cost by spending fewer tokens through the effort parameter. One approach gives a lower price floor; the other gives finer control on a single quality profile.
Q: Can thinking be disabled on Claude Opus 5?
A: Yes, but only at high effort or below. Setting thinking to disabled with xhigh or max effort returns a 400 error. If you want thinking disabled, keep it at high effort or remove the thinking field entirely.
Discussion
0 comments