Claude Opus 5: specs, effort levels, and when to choose it
Anthropic launched Opus 5 with near-Fable 5 intelligence at half the cost — default thinking and five effort levels. A practical guide to specs, comparisons, and when to escalate from Sonnet.
What makes Opus 5 stand out?
In July 2026, Anthropic launched Claude Opus 5 — not an incremental update, but a leap in complex coding and long-horizon agents. The model approaches Fable 5 intelligence on hard tasks at significantly lower cost. If you already pay for Opus 4.8, upgrading delivers higher performance at the same pricing.
Production specs that matter
Context
1M tokens
Max and default — suited for large documents and big projects.
Output
128K tokens
Max output — with thinking enabled by default on every turn.
Pricing
$5 / $25
Input / output per 1M tokens — same as Opus 4.8.
Knowledge
May 2026
Updated cutoff — API id: claude-opus-5.
Thinking
On by default
The model decides when and how much to think — no manual toggle needed.
Fast mode
2.5× faster
Higher speed at 2× price — for high-value tasks only.
Effort levels — the defining 2026 trend
Opus 5 expands the effort ladder to five grades: low, medium, high (default), xhigh, and max. The same model runs at low effort for routine work or maximum effort for critical tasks — with a bigger quality gap than any prior generation.
low / medium
Routine tasks — summaries, classification, first drafts.
high (default)
Start here — tune before escalating.
xhigh
Complex coding — thinking required, max_tokens 64K+.
max
Maximum quality — near-Fable 5 for critical work.
At xhigh and max, thinking cannot be disabled. For short replies, request explicitly in the prompt.
Opus 5 vs Opus 4.8
Opus 5 vs Opus 4.8
| Dimension | Opus 4.8 | Opus 5 |
|---|---|---|
| Pricing | $5 / $25 | $5 / $25 (unchanged) |
| Context window | 1M tokens | 1M tokens |
| Thinking | Adaptive | Enabled by default |
| Effort levels | Up to xhigh | Up to max |
| Coding | Strong | 2×+ on Frontier-Bench |
| Fast mode | Available | 2.5× speed |
Pricing in the Claude lineup
Input / 1M tokens
Output / 1M tokens
Opus 5 vs GPT-5.6 Sol
Opus 5 vs GPT-5.6 Sol
| Dimension | Opus 5 | GPT Sol |
|---|---|---|
| Output pricing | $25/1M | $30/1M |
| Context window | 1M tokens | ~1.05M tokens |
| Opus 5 strength | Long-horizon agents + wider effort ladder | — |
| Sol strength | — | Codex + parallel Sol Ultra |
| Best for | Enterprise production coding | OpenAI agents + computer use |
When should you pick Opus 5?
When to pick Opus 5 over Sonnet or Fable
Opus 5
System refactors, hour-long coding agents, large documents at high accuracy.
Opus 5 at max
Within 0.5% of Fable 5 performance at half the task cost.
Fable 5
Sensitive research, strategic decisions, advanced cybersecurity.
Start with Sonnet
Escalate to Opus 5 only when cheaper tiers fail.
Practical tips for product teams
Practical tips for product teams
- 1
Don't use Opus 5 for every request
Start with Sonnet or Haiku — escalate only when cheaper tiers fail.
- 2
Tune effort before switching models
Many tasks resolve at high instead of max — saves tokens significantly.
- 3
Monitor cost by effort level
Dashboard showing spend per level (high / xhigh / max).
- 4
Use Fast mode carefully
For high-value quick tasks — not daily high volume.
- 5
Enable prompt caching
Up to 90% off repeated inputs and 50% off batch processing.
The takeaway
Opus 5 makes the Opus tier the default for hard work in 2026: near-Fable 5 performance, stable pricing, and finer control via effort levels. For the full Claude lineup, see the [Claude models 2026 comparison](/en/insights/claude-models-2026).