What Is Claude Sonnet 5? Pricing, Capabilities & API Guide (2026)
Anthropic replaced Sonnet 4.6 on every free and paid plan today. The new model doubles max output, adds native multi-agent coordination, and launches with an introductory price that undercuts the old model through August 31. Here is what changed, what it costs, and who should care.
intro through Aug 31
tokens
2x Sonnet 4.6
Dev Team, self-verify, injection defense
What Is Claude Sonnet 5?
Claude Sonnet 5 is Anthropic's everyday-tier model, released June 30, 2026. It replaced Sonnet 4.6 as the default for Free plan users, Pro plan subscribers, claude.ai, and Claude Code on the same day it launched. In Anthropic's model hierarchy, Sonnet sits between Haiku (fast, cheap) and Fable 5 (frontier, expensive). Sonnet 5 is the model most people hit when they open claude.ai or call claude-sonnet-5 through the API. For the full timeline of how the family evolved from Opus 4.5 to this release, see the Claude model lineage.
Put governance around how your team uses AI. The AI Acceptable Use Policy: a deploy-ready template that sets the rules for AI use.
Your purchase helps keep our hubs free to read.
Anthropic structures Claude 5 as two publicly available tiers: Sonnet 5 (everyday default) and Fable 5 (frontier, $10/$50 per MTok). A third model, Mythos 5, was released alongside Fable 5 but is restricted to Project Glasswing partners and is not available to the general public or through standard API access.
Haiku 4.5 ($1/$5/MTok) → Sonnet 5 ($2/$10 intro / $3/$15 std) → Opus 4.8 ($5/$25) → Fable 5 ($10/$50). Sonnet 5 is the new performance sweet spot: more capable than Opus 4.8 was two years ago, at roughly half Opus 4.8's current price.
The launch date matters for one practical reason: Sonnet 5 uses an updated tokenizer, inherited from the Opus 4.7 generation, that consumes 1.0x to 1.35x more tokens per equivalent text compared to Sonnet 4.6. The introductory pricing ($2/$10) partially compensates for this through August 31, 2026, after which the rate rises to the standard $3/$15. Any application doing heavy text processing should benchmark token consumption on Sonnet 5 before the rate change.
How It Works: Dev Team Mode, Self-Verification, Injection Resilience
Three capability additions distinguish Sonnet 5 from Sonnet 4.6. None of them require special configuration, but all three affect how you'd architect an agent or evaluate suitability for a task.
Dev Team Mode
Dev Team mode is native multi-agent orchestration inside a single Sonnet 5 call. Instead of a single agent working linearly through a task, Sonnet 5 can split work among specialized sub-agents that run in parallel: a planner, one or more coders, a reviewer, and a tester. The model coordinates result synthesis automatically. For a software project with clearly separable components, Dev Team mode reduces wall-clock time by handling multiple workstreams concurrently rather than sequentially.
This is distinct from manually chaining Claude API calls with an external orchestrator. Dev Team mode does the decomposition and coordination internally, which means the planning overhead is reduced and context passes efficiently between the sub-agents without serializing through your application layer.
Self-Verification for Code
Self-verification is Sonnet 5's ability to close the bug-fix loop in one pass: it writes a reproducing test, implements the fix, re-runs the test, and confirms the fix holds before returning the result. On Sonnet 4.6 and earlier, this required multiple turns or explicit orchestration logic from the caller. In Sonnet 5 it happens in a single API call when the task context makes it appropriate.
The practical effect: fewer back-and-forth iterations on bug fixes, and fewer cases where the model returns a "fix" that doesn't address the root cause. The test acts as a forcing function the model uses to validate its own output before the caller ever sees it.
Prompt Injection Resilience
Prompt injection resilience received significant improvement in Sonnet 5. In agentic workflows, the model encounters user-controlled content, document text, web pages, and tool outputs that could contain adversarial instructions. Sonnet 5 is substantially better at distinguishing its system instructions from injection attempts in that content. For production applications where the model processes external data, this reduces the attack surface without requiring the caller to implement additional filtering layers.
Sonnet 5 uses an updated tokenizer that encodes the same text as 1.0x to 1.35x more tokens compared to Sonnet 4.6. The introductory pricing compensates for most workloads through August 31, 2026. Heavy-throughput applications should benchmark token consumption before the price rises to $3/$15 standard.
Pricing
Pricing current as of June 30, 2026. Introductory rates expire August 31, 2026. Verify at anthropic.com/pricing before committing to a budget. For a tier-by-tier pricing and API reference, see the Claude Sonnet 5 pricing and API guide.
| Model | Input /MTok | Output /MTok | Cache Write | Cache Read | Notes |
|---|---|---|---|---|---|
| Sonnet 5 (intro) NEW | $2.00 | $10.00 | $3.75 | $0.30 | Through Aug 31, 2026 |
| Sonnet 5 (standard) | $3.00 | $15.00 | $3.75 | $0.30 | From Sep 1, 2026 |
| Opus 4.8 | $5.00 | $25.00 | $6.25 | $0.50 | Strong reasoning, agentic |
| Fable 5 | $10.00 | $50.00 | $12.50 | $1.00 | Frontier, 95.5% SWE-bench |
| Sonnet 4.6 (legacy) | $3.00 | $15.00 | -- | -- | Same price as Sonnet 5 standard after Aug 31 |
| Haiku 4.5 | $1.00 | $5.00 | -- | -- | Fast, cost-efficient |
Batch API discount: 50% off all models. Context windows: 1M tokens across all listed tiers. Max output: 128K for Sonnet 5, Fable 5, Opus 4.8; 64K for Sonnet 4.6.
Sonnet 5's updated tokenizer may encode the same prompt as 1.0x to 1.35x more tokens than Sonnet 4.6. For a typical 1,000-token Sonnet 4.6 prompt, plan for up to 1,350 tokens on Sonnet 5. At intro pricing ($2/MTok vs $3/MTok on Sonnet 4.6), the cost impact is minimal until August 31. After that, budget for the tokenizer delta when switching applications from Sonnet 4.6 to Sonnet 5 standard.
Claude Sonnet 5 vs Sonnet 4.6
| Capability | Sonnet 4.6 | Sonnet 5 UPGRADE |
|---|---|---|
| Max output tokens | 64K | 128K (2x) |
| Context window | 1M tokens | 1M tokens (unchanged) |
| Multi-agent coordination | External orchestration only | Dev Team mode (native) |
| Bug-fix loop | Multiple turns required | Self-verification (single pass) |
| Prompt injection resilience | Baseline | Significantly improved |
| Tokenizer | Pre-Opus-4.7 generation | Updated (1.0x-1.35x more tokens) |
| API pricing (input) | $3.00/MTok | $2.00/MTok intro (through Aug 31) |
| Default on Free plan | Yes (until Jun 30) | Yes (from Jun 30) |
Max output tokens. Context window (1M) is identical across both. Doubling max output lets Sonnet 5 handle longer generations, larger code files, and extended document summaries without truncation.
Who Should Use Claude Sonnet 5
Claude Code + Sonnet 5
Sonnet 5 is now the default model for Claude Code as of June 30, 2026. Claude Code is the agentic coding CLI that runs directly in your terminal, reads your codebase, executes commands, and iterates on fixes autonomously. On Sonnet 5, Claude Code benefits from all three new agentic capabilities: Dev Team mode for parallel workstreams, self-verification to close bug-fix loops without prompting, and injection resilience to handle external file content more safely.
For users on the Pro plan ($20/month), Claude Code access is included and will now use Sonnet 5 by default. Team Premium ($150/user/month) includes Claude Code terminal integration with additional usage. Developers who need Fable 5 for Claude Code -- for example, when working on the most complex agentic pipelines or long-horizon autonomous tasks -- can specify claude-fable-5 explicitly in their configuration.
If you use Claude Code on any paid plan, Sonnet 5 is now active automatically. You don't need to update the CLI, change configuration, or switch model IDs. The switch happened on the backend when Sonnet 5 launched.
One migration task worth running: if your team has hard-coded claude-sonnet-4-6 in any API call, application config, or system prompt, you're now using the legacy model. claude-sonnet-5 is the correct model ID for Sonnet 5. For teams that want the latest Sonnet automatically without updating code on each release, the alias claude-sonnet-latest will always resolve to the current Sonnet generation.
Limitations
No published independent benchmarks exist for Sonnet 5 at launch. All capability claims about Dev Team mode, self-verification quality, and injection resilience are Anthropic-reported. This is standard for day-one model launches -- third-party evaluations (Artificial Analysis, METR, LMSYS) require time to run. Budget a few weeks before treating Sonnet 5 benchmark claims as independently verified. For how Sonnet 5 lines up against OpenAI's current flagship on reasoning and coding, see Claude Sonnet 5 vs GPT-5.6.
The tokenizer change is the practical near-term risk. Applications that sized budgets or rate-limit logic against Sonnet 4.6 token counts may need adjustments. The 1.0x to 1.35x range is wide: for casual prompts the impact is minor, but for high-density technical content or code-heavy inputs, token consumption is closer to the 1.35x end. Run representative samples through the API before relying on prior cost estimates.
Sonnet 5 is not a replacement for Fable 5 in frontier use cases. If your workflow requires Fable 5's 95.5% SWE-bench Verified score, multi-day autonomous task runs with persistent file-based memory, or the highest available benchmark performance on agentic coding, Sonnet 5 is not the equivalent. It is the default model, not the ceiling.
Audio and video processing remain outside Claude's capabilities entirely. Sonnet 5 handles text and images. No native image generation. No audio transcription. No video analysis. If those capabilities matter for your use case, a different tool is still required alongside Claude. For the rest of the Claude guides, reviews, and comparisons, browse the Claude hub.
The $2/$10 introductory rate expires August 31, 2026. From September 1, pricing rises to $3/$15 standard, matching the old Sonnet 4.6 rate but with the updated tokenizer consumption factored in. Applications running at scale should plan for this before the rate change, not after.