Models: Opus, Sonnet, Haiku

Last verified: 5 May 2026 | Applies to: All plans

In 30 seconds

Claude comes in three tiers: Opus (most capable, highest cost), Sonnet 4.6 (balanced performance and speed, best value for most operator tasks), and Haiku 4.5 (fastest, lowest cost). Opus 4.7 is the new flagship model (launched 16 April 2026), with major gains in coding and vision. Most operators should still default to Sonnet 4.6 for everyday work, as it beats Opus on real-world office tasks. Haiku is for high-volume, simple tasks where speed matters more than depth.

The models

	Opus 4.7	Opus 4.6	Sonnet 4.6	Haiku 4.5
Capability	New flagship. Highest on coding, vision, and complex reasoning	Strong on complex coding and research	Highest on real-world office tasks (drafting, analysis, reporting)	Basic, good for simple, fast tasks
Speed	Slowest	Slow	Moderate	Fastest
Cost (API)	$5/1M input, $25/1M output	$5/1M input, $25/1M output	Lower	Lowest
Context window	1M tokens (standard pricing)	1M tokens (standard pricing)	1M tokens (standard pricing)	200K tokens
Vision	3.75MP (3x Opus 4.6)	1.15MP	Standard	Standard
Best for	Complex coding, vision-heavy work, deep research, tasks where 4.6 falls short	Complex coding, deep research, edge cases where Sonnet falls short	Everyday work, drafting, data processing, financial analysis, most operator tasks	Quick lookups, simple formatting, high-volume repetitive tasks

Opus 4.7 highlights (launched 16 April 2026): 87.6% SWE-bench Verified (up from 80.8% on 4.6), 64.3% SWE-bench Pro (up from 53.4%), 3.75MP vision resolution (3x the 1.15MP on 4.6), new tokeniser that fits roughly 35% more tokens from the same input, same API pricing as Opus 4.6, and a new “xhigh” effort level. Task Budgets is available in beta. Anthropic notes that Opus 4.7 is “less broadly capable” than Mythos Preview, meaning it excels at coding and vision but may not lead across every benchmark. Available on all paid plans.

Opus 4.7 long-context regression. Long-context recall (MRCR) dropped from 78.3% on Opus 4.6 to 32.2% on Opus 4.7. If you work with long legal or financial documents, test your results against Opus 4.6 before committing to 4.7 for these workflows.

Competitive context (April 2026). OpenAI launched GPT-5.5 on 23 April 2026. Claude Opus 4.7 leads on 6 of 10 shared benchmarks including coding and reasoning. GPT-5.5 leads on agentic terminal tasks and web browsing. For operators, Claude remains the stronger choice for deep document work and analysis.

When to use each model

graph TD
    A{What's the task?} -->|Simple and repetitive| C[Haiku 4.5]
    A -->|Complex coding or deep research| E[Try Sonnet 4.6 first]
    A -->|Everything else| F[Sonnet 4.6]
    E --> G{Output good enough?}
    G -->|Keep using Sonnet| F
    G -->|No| H[Opus 4.7]

Default to Sonnet 4.6 for:

• Drafting emails, reports, and documents
• Analysing spreadsheets and data
• Financial analysis and reporting
• Using plugins for standard workflows
• Most Cowork tasks
• Day-to-day chat conversations

Sonnet 4.6 scores higher than Opus 4.6 on real-world office task benchmarks (GDPval-AA) and agentic financial analysis, at one-fifth the cost per message. For most operator workflows, Sonnet 4.6 is now the best model for the money.

Switch to Opus 4.7 (recommended) or Opus 4.6 for:

• Complex coding and software development
• Deep research requiring extensive reasoning chains
• Processing very long documents (contracts, regulatory filings)
• Vision-heavy work with high-resolution images (Opus 4.7 handles 3.75MP)
• Edge cases where Sonnet’s output isn’t quite good enough

Opus 4.7 is the recommended Opus model. Opus 4.6 remains available if you prefer it or encounter a task where 4.6 performs better for your specific use case.

Adaptive thinking (Opus 4.7, Opus 4.6, and Sonnet 4.6): These models use adaptive thinking. Claude dynamically decides when and how much to think based on task complexity, using effort levels (low, medium, high, max, plus “xhigh” on Opus 4.7). Simple questions get a fast, direct response. Complex analysis triggers deeper reasoning automatically. This replaces the binary extended thinking toggle used by earlier models. You don’t need to manually switch thinking on. Claude handles this automatically. Haiku 4.5 does not have adaptive thinking.

Default effort was quietly lowered

In early March 2026, Anthropic lowered the default effort level from “high” to “medium”. If responses feel shorter or less thorough than they used to, set effort to “high” explicitly in your system prompt, or use the /effort command in Claude Code.

Use Haiku 4.5 for:

• Quick questions with straightforward answers
• Simple formatting or conversion tasks
• High-volume tasks where you need speed (e.g., categorising many items)
• Claude in Chrome on the Pro plan (Haiku is the only option)

Model access by plan

Plan	Chat models	Chrome models	Notes
Free	Sonnet 4.6, Haiku 4.5	—	No Chrome access
Pro	Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5	Haiku 4.5 only	Chrome limited to Haiku
Max	Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5	Opus 4.6, Sonnet 4.5, Haiku 4.5	Chrome not yet updated for Sonnet 4.6 or Opus 4.7
Team	Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5	Opus 4.6, Sonnet 4.5, Haiku 4.5	Chrome not yet updated for Sonnet 4.6 or Opus 4.7
Enterprise	Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5	Opus 4.6, Sonnet 4.5, Haiku 4.5	Chrome not yet updated for Sonnet 4.6 or Opus 4.7

Context windows explained

A context window is how much text Claude can process in a single conversation. 200K tokens is roughly 150,000 words, enough for most business documents.

1M token context is now generally available at standard pricing for both Opus 4.6 and Sonnet 4.6 (as of 13 March 2026). No beta header or special access required. The media limit has also increased from 100 to 600 images or PDF pages per request. Available on Max, Team, Enterprise, and API. Pro users can access 1M context in Claude Code via /extra-usage. 1M tokens is roughly 750,000 words, enough for an entire quarter’s financials or a complete contract set with amendments in a single conversation.

Sonnet 4.6 knowledge cutoffs: Reliable knowledge through August 2025. Training data through January 2026.

Retired, deprecated, and upcoming changes

• Sonnet 3.7 and Haiku 3.5 were retired on 19 February 2026. API requests to these models now return errors.
• Haiku 3 was retired on 19 April 2026. API calls now return errors. The replacement is Haiku 4.5, which costs roughly 4x more per token.
• Sonnet 4 and Opus 4 were deprecated on 14 April 2026 and will be retired from the API on 15 June 2026. Migrate to Sonnet 4.6 or Opus 4.7/4.6 before that date.
• Sonnet 4.5 and Sonnet 4 lose 1M context on 30 April 2026. That deadline is days away. The 1M context beta for these older models is being retired. If you still rely on 1M context with Sonnet 4.5 or 4.0, migrate to Sonnet 4.6 or Opus 4.6/4.7 immediately. Those models have 1M context at standard pricing.

Mythos: next-gen model leaked, not released

On 26 March 2026, a CMS misconfiguration exposed details of a next-gen model codenamed “Capybara” (Mythos). It’s described as larger and more intelligent than Opus, with dramatically higher benchmark scores and unprecedented cybersecurity capabilities. Mythos is not publicly available. It is in early access only, with no release date announced. There is no pricing, no plan availability, and no way to use it yet. We’ll cover it properly once Anthropic makes an official announcement. See the Changelog for details on the leak.

How operators think about this

“I just want it to work well.” Use Sonnet 4.6. It handles 90% of operator tasks and actually outperforms Opus on everyday business work like drafting, analysis, and reporting.

“I need the absolute best for a complex task.” Try Sonnet 4.6 first. It may surprise you. Switch to Opus 4.7 if you’re doing deep coding, extended research, or if Sonnet’s output isn’t quite right. You can change models per conversation.

“My team is burning through usage too fast.” Have team members default to Haiku for simple tasks and Sonnet for real work. Reserve Opus for tasks that genuinely need it. See Cost Management.

“Claude in Chrome isn’t smart enough on my Pro plan.” That’s because Pro limits Chrome to Haiku 4.5. If you need Chrome for complex automation, upgrade to Max for Sonnet/Opus access in the browser.

What to watch out for

• Sonnet 4.6 beats Opus on most office tasks. Don’t assume Opus is always better. For everyday operator work, Sonnet 4.6 actually scores higher on office benchmarks and uses far less of your plan’s usage allocation.
• Model names change. Anthropic releases new model versions periodically. The version numbers (4.6, 4.5) will change, but the tier structure (Opus > Sonnet > Haiku) has remained stable.
• Usage varies by model. Opus consumes more of your plan’s usage allocation per message than Sonnet or Haiku. This matters on plans with usage limits.
• Chrome hasn’t got Sonnet 4.6 yet. Claude in Chrome still offers Opus 4.6, Sonnet 4.5, and Haiku 4.5. Pro is limited to Haiku 4.5 only.

Related

• Choosing a Plan: model access varies by plan
• Claude in Chrome: Pro users are limited to Haiku 4.5
• Cost Management: optimising model usage across a team

---

Something wrong or outdated? Let us know →

Get weekly workflows: subscribe to the newsletter.