Anthropic released Claude Sonnet 5 on June 30, 2026 — and it closes a gap that has been widening for months. For the past year, the most capable agentic performance lived in Opus-class models. Sonnet 5 brings that capability down to Sonnet pricing.
What Changed
Sonnet 5 is a direct successor to Sonnet 4.6 and targets four core dimensions: reasoning, tool use, coding, and knowledge work. Anthropic’s characterization is deliberate: this is the most agentic Sonnet yet. It can make multi-step plans, operate browsers and terminals, and run autonomously at a level that previously required Opus-class compute.
The historical context matters here. Claude Sonnet 3.5, 3.6, and 3.7 were the models that first demonstrated impressive agentic skills — coding agents, tool-chaining, multi-step workflows. After that, the clearest capability gains shifted to Opus. Sonnet 5 is Anthropic’s move to close that regression.
Early access partners described a consistent pattern: Sonnet 5 finishes complex agentic tasks where Sonnet 4.6 stopped short. It checks its own output without being asked. It recovers from tool failures without giving up.
Pricing
Sonnet 5 launches with introductory pricing through August 31, 2026:
| Model | Input | Output |
|---|---|---|
| Sonnet 5 (intro, through Aug 31) | $2/MTok | $10/MTok |
| Sonnet 5 (standard, from Sep 1) | $3/MTok | $15/MTok |
| Opus 4.8 (for reference) | $5/MTok | $25/MTok |
| Sonnet 4.6 (predecessor) | $3/MTok | $15/MTok |
At introductory pricing, Sonnet 5 is cheaper than Sonnet 4.6 was at standard rates — and meaningfully cheaper than Opus 4.8 at any price point.
One important caveat: Sonnet 5 uses a new tokenizer. The same input that was processed by Sonnet 4.6 will produce roughly 1.0–1.35× more tokens on Sonnet 5, depending on content type. The introductory pricing is set to be roughly cost-neutral during the transition, but re-benchmark your actual workloads before September.
Benchmarks
Anthropic published cost-performance curves at different effort levels for two evaluations:
BrowseComp (agentic web search): Sonnet 5 at higher effort levels matches Opus 4.8 on some tasks. Sonnet 4.6 fell well short of Opus 4.8 on this evaluation.
OSWorld-Verified (agentic computer use): Sonnet 4.6 scored 78.5% (updated baseline). Sonnet 5 is a strict improvement at every effort level.
The practical reading: Sonnet 5 at high effort overlaps with Opus 4.8 at medium effort on agentic tasks, at a substantially lower token cost.
API Access
The model ID is claude-sonnet-5. It is available on the Claude Platform (native, AWS, Azure/Foundry, Google Vertex), Claude Code, and is the default model on Free and Pro plans starting today.
import anthropic
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
messages=[
{"role": "user", "content": "Plan and execute a multi-step analysis of this codebase."}
]
)Rate limits have been increased across all tiers — Chat, Cowork, Claude Code, and the API — to accommodate the higher token usage that comes with higher effort levels.
Availability
Sonnet 5 is available today across all Claude plans:
- Free and Pro: Default model
- Max, Team, Enterprise: Available alongside other models
- Claude Code: Available today
- Claude Platform API: Available with model ID
claude-sonnet-5
Safety Profile
Pre-deployment evaluations found Sonnet 5 is overall safer than Sonnet 4.6:
- Better prompt injection resistance — stronger at detecting and refusing malicious requests in agentic contexts
- Lower hallucination rate than Sonnet 4.6
- Lower sycophancy rate than Sonnet 4.6
- Cyber safeguards enabled by default — same safeguards present in Opus 4.7 and 4.8, less strict than Fable 5
On cybersecurity tasks specifically: Sonnet 5 shows substantially weaker exploit-development ability than Opus 4.8 or Fable 5. Neither Sonnet 4.6 nor Sonnet 5 could develop a working exploit in Anthropic’s Firefox 147 evaluation — Sonnet 5 showed a slightly higher partial success rate, likely from general intelligence improvements rather than specific training. For cybersecurity work requiring reduced guardrails, Anthropic recommends Opus 4.8.
My Take
Upgrade your agentic workloads. If you are running Claude Code, multi-step tool-use pipelines, or any workflow that previously required Opus 4.8 to complete reliably, Sonnet 5 at introductory pricing is the obvious move. The cost-performance curve has shifted substantially.
Mind the tokenizer. The 1.0–1.35× token multiplier is real. If you have tight latency SLAs or token budget constraints, measure before migrating. Prompt caching behavior will also shift — cached prefix sizes may differ between Sonnet 4.6 and 5.
The September cliff matters. Introductory pricing ends August 31. At standard pricing, Sonnet 5 ($3/$15) costs the same as Sonnet 4.6 did, but the tokenizer change means your actual spend will be higher for the same prompts. Factor that into your September cost projections.
For my own deployments — Claude Code, automated blog pipelines, and research agents — I am switching to Sonnet 5 today. The agentic completion improvement is not marginal. Partners described it as “finishes what Sonnet 4.6 started.” That is worth the tokenizer adjustment.
Migration Checklist
Before switching claude-sonnet-5 into production:
- Re-benchmark your actual prompt token counts with the new tokenizer
- Test prompt cache hit rates — cached prefix boundaries may shift
- Verify your agentic workflows complete end-to-end (not just first steps)
- Review tool-use error handling — Sonnet 5 recovers more aggressively, which may change your retry logic
- Update cost projections for September standard pricing
