# Claude Sonnet 5 closes the gap to Opus without the Opus bill

> Sonnet 5 is the default on Free and Pro at $2/$10 per million tokens - close to Opus 4.8 on Anthropic's launch benchmarks at 40% of the Opus input price.

URL: https://bartlomiejkrupa.dev/articles/claude-sonnet-5/
Published: 2026-06-30
Updated: 2026-10-07
Tags: claude, sonnet, opus, pricing, agents, model-selection, fable
Summary: Sonnet 5 is the default on Free and Pro at $2/$10 per million tokens - close to Opus 4.8 on Anthropic's launch benchmarks at 40% of the Opus input price.

> **Update (2026-10-07):** Sonnet 5 is now a legacy model. Its successor, **Claude Sonnet 5.5** (`claude-sonnet-5-5`), keeps the $2 / $10 price, and the current Opus is Opus 5.5 at $4 / $20. The role advice below still holds with the 5.5 IDs - see the [tier guide](/articles/claude-model-tier-comparison) and the [5.5 migration checklist](/articles/claude-5-5-migration).

Anthropic shipped **Claude Sonnet 5** (`claude-sonnet-5`) on 2026-06-30 as the most agentic Sonnet tier yet - positioned close to **Opus** on agentic search and computer-use benchmarks, at **$2 / $10** per million input/output tokens (announced as an intro price; the planned move to $3 / $15 did not happen). It is the default on Free and Pro, available in Claude Code and on the Claude Platform. For agent workflows, treat it as the upgraded daily driver and sub-agent worker; keep [Opus 5](/articles/claude-opus-5) for orchestration and tasks where you need the highest accuracy at High or XHigh effort.

## Definitions

**Sonnet 5** - Anthropic's mid-tier model (`claude-sonnet-5`). Replaces Sonnet 4.6 as the default Sonnet class. Anthropic reports strict improvement over Sonnet 4.6 on BrowseComp (agentic search) and OSWorld-Verified (computer use), with cost-performance curves that overlap Opus's range at lower per-token price.

**Effort level** - a Sonnet/Opus parameter that trades token spend for accuracy. Sonnet 5 defaults to `high` on the API, so Medium is a setting you must choose. Higher effort spends more tokens per answer - Anthropic raised rate limits at launch to absorb exactly that - so the bill advantage shrinks at the top of the curve.

**Cyber safeguards** - real-time blocks on dangerous cybersecurity usage. Enabled by default on Sonnet 5 (less strict than Fable 5). For cyber work that needs reduced guardrails, Anthropic's launch post points at Opus 4.8 - Sonnet 5 was not trained for those tasks, so lowering its safeguards buys nothing.

## What shipped

| | Sonnet 5 | Opus 5 |
| --- | --- | --- |
| API ID | `claude-sonnet-5` | `claude-opus-5` |
| Input / output ($/M tokens) | $2 / $10 | $5 / $25 |
| Default on claude.ai | Free + Pro | Max (strongest on Pro) |
| vs Sonnet 4.6 | Strict improvement on agentic benchmarks; lower misalignment and hallucination rates; tokenizer caveat (below) | Higher peak accuracy on long-horizon agentic work |
| Cyber capability | Substantially weaker than Opus on dangerous exploit tasks; safeguards on by default | Stronger; Anthropic's Sonnet 5 launch recommended Opus (then 4.8) for cyber work needing reduced guardrails |

**Tokenizer change:** Sonnet 5 uses an updated tokenizer (similar to Opus 4.7). The same text can map to **~1.0–1.35×** more tokens depending on content type.

**Availability:** Claude Code, Claude Platform API, and all claude.ai plans from launch day. Anthropic also raised rate limits across Chat, Cowork, Claude Code, and the Platform for higher-effort runs.

Sources: [Introducing Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5). Orchestrator ID updated for [Opus 5](/articles/claude-opus-5) (2026-07-24). Pricing note (2026-09-30): the $2 / $10 rate did not expire on 2026-08-31, and the successor Sonnet 5.5 (`claude-sonnet-5-5`) launched at the same price - current tiers in [the tier guide](/articles/claude-model-tier-comparison).

## Where Sonnet 5 sits in the stack

The role pattern from [pick the right Claude tier](/articles/claude-model-tier-comparison) still holds - Fable 5 is the frontier rung, Opus 5 is the orchestrator, Sonnet 5 is the worker:

| Role | Tier | Why (post–Opus 5) |
| --- | --- | --- |
| Orchestrator | [Opus 5](/articles/claude-opus-5) | Longest horizon, highest autonomy below Fable, best top-end accuracy on agentic evals at $5/$25 |
| Default worker | **Sonnet 5** | Near-Opus agentic quality at 40% of Opus 5's input cost ($2 vs $5) |
| Parallel worker | Haiku 4.5 | Still the cheapest fan-out tier until Haiku gets its own bump |

Early-access partners quoted on Anthropic's Sonnet 5 launch post describe end-to-end task completion, unprompted self-checking, brownfield debugging, and fewer steps to the same outcome - the behaviors that matter for sub-agents spawned under an Opus orchestrator. See [subagent context isolation](/notes/subagent-context-isolation) for why the worker tier's quality matters more than the orchestrator's benchmark delta.

Encoded for launchers and config:

```bash
ORCHESTRATOR=claude-opus-5     # long-horizon, max accuracy at high effort
DEFAULT=claude-sonnet-5        # daily driver + sub-agent default (was sonnet-4-6)
WORKER=claude-haiku-4-5        # parallel scoped tasks
```

## Reading the cost-performance curve

Anthropic's launch charts plot **cost vs accuracy** at different effort levels on BrowseComp and OSWorld-Verified. Three takeaways for the bill:

1. **Sonnet 5 strictly beats Sonnet 4.6** at every effort level on those curves - a real upgrade, not a rename.
2. **Sonnet 5 and Opus cover one range** - Sonnet is cheaper at comparable effort; Opus still wins on peak accuracy.
3. **High and XHigh effort on Sonnet 5 spend more tokens per point of accuracy** - at the top of the curve, total cost can approach or exceed Opus for the same task quality.

Practical rule: run Sonnet 5 at **Medium effort** (or lower) for the worker and default roles. Escalate to Opus 5 - not Sonnet 5 at XHigh - when the task needs frontier accuracy and you are willing to pay for it.

## When to stay on Sonnet 5 vs reach for Opus

| Situation | Tier |
| --- | --- |
| Interactive coding, document Q&A, most Claude Code sessions | Sonnet 5 |
| Sub-agents under an Opus orchestrator | Sonnet 5 |
| Long-horizon migration, hundreds of tool calls, plan must not drift | Opus 5 |
| Cybersecurity work that needs reduced guardrails | Opus (Anthropic named Opus 4.8 at launch) |
| You are burning Pro limits in ~30 minutes on Opus | Sonnet 5 - lower per-token rate extends the same subscription |

Sonnet 5 is **not** a Fable replacement. Anthropic did not train it for dangerous cyber tasks; it scores substantially below Opus and Fable 5 on exploit-development evaluations and ships with safeguards on by default.

## Verify the effort rule on your own workload

The Medium-effort rule above is the one claim here worth testing rather than taking on faith, because it decides your bill. Run the same task three times - Sonnet 5 at Medium, Sonnet 5 at XHigh, Opus 5 at High - and log total tokens and task success for each.

The failure the rule predicts is specific: Sonnet 5 at XHigh spends materially more tokens without closing the accuracy gap, so it lands at or above Opus 5's cost for worse results. If you see that crossover on your tasks, the escalation path is the model, not the dial. If you don't, keep the cheaper setting - the curves are Anthropic's launch measurements on their benchmarks, not a guarantee about your workload.

## Bottom line

Sonnet 5 is the default worker tier: agentic enough to shrink the gap to Opus, priced to run all day. Default to Sonnet 5, orchestrate with Opus 5, fan out on Haiku 4.5, and match **effort level to the role** - Medium Sonnet for volume, High Opus when accuracy is the bottleneck. Escalate the tier, not the effort dial.

Sources:
- Introducing Claude Sonnet 5: https://www.anthropic.com/news/claude-sonnet-5
- Claude Sonnet 5: https://platform.claude.com/docs/en/models/sonnet-5/overview
- Models overview: https://platform.claude.com/docs/en/models/overview