Claude Opus vs Sonnet vs Haiku (2026): Which to Use When

TL;DR

Use Claude Haiku 5.5 ($0.10/$0.50 per million input/output tokens) for triage and high-volume classification, Claude Sonnet 5.5 ($2/$10) for everyday building, and Claude Opus 5.5 ($4/$20) for review and long-running agentic coding. Opus costs twice Sonnet per token, and Sonnet costs 20 times Haiku 5.5. Claude Fable 5.1 ($10/$50) sits above Opus as the escalation step.

Prices checked October 23, 2026 against Anthropic's pricing page and models overview. Claude API list prices; Bedrock, Google Cloud and enterprise contracts can differ.

Anthropic's current lineup is now all on the 5.5 generation below Fable: Opus 5.5 (September 22, 2026), Sonnet 5.5 (September 28, 2026) and Haiku 5.5 (October 7, 2026), plus Fable 5.1 at the top. Haiku 5.5 is the big change: it lists at a tenth of Haiku 4.5's per-token price and moves to a 1M context window. For Opus release details and breaking changes, read our Claude Opus 5.5 release breakdown.

Opus is Anthropic's model "for long-running agentic coding and knowledge work", with moderate latency. Sonnet is "the best combination of speed and intelligence", with fast latency. Haiku is "for high-volume, latency-sensitive tasks such as classification, extraction, and routing", with the fastest latency and the lowest price. Those are Anthropic's own descriptions, and they line up with the prices.

One wrinkle before you pick: Anthropic's own selection guide says "most workloads start with Claude Opus 5.5." Our routing rule below is the cost-aware version of that advice. Haiku triages, Sonnet builds, Opus reviews, and you use your own evals to decide how much work each tier takes. This guide shows how that maps to real tasks and how we wire the routing inside production agents at AY Automate.

What changed in September and October 2026

Anthropic replaced all three tiers in five weeks, so older Opus vs Sonnet vs Haiku guides are now wrong on price and specs. The often-quoted $15/$75 Opus price belongs to Opus 4 and 4.1, which are retired on Anthropic's API.

Details from Anthropic's pricing page and model pages:

• Opus 5.5 (released September 22, 2026): $4 input, $20 output, $0.20 cache reads, $2/$10 in batch. Opus 5 and Opus 4.8 remain available at $5/$25.
• Sonnet 5.5 (released September 28, 2026): $2 input, $10 output, the same as Sonnet 5, with cache reads cut from $0.20 to $0.10. It uses the same tokenizer as Sonnet 5, so token counts don't change.
• Haiku 5.5 (released October 7, 2026): $0.10 input and $0.50 output for prompts up to 100,000 tokens, $0.50/$2.50 above that. Haiku 4.5 is still listed at $1/$5. Haiku 5.5 uses the newer tokenizer, so the same text counts as about 30% more tokens than on Haiku 4.5.
• Fable 5.1 (released September 1, 2026): same $10/$50 as Fable 5, with cache reads cut to $0.25.
• Opus 4 and Opus 4.1 ($15/$75): retired on Anthropic's API. Opus 4.1 remains on Bedrock and Google Cloud, Opus 4 on Google Cloud only.

At a glance: the comparison table

The table below copies Anthropic's pricing page and models overview as printed on October 23, 2026. Prices are per million tokens on the Claude API. Haiku 5.5, Sonnet 5.5 and Opus 5.5 form the classic three-tier ladder; Fable 5.1 is included because Anthropic lists it above Opus.

Notes on the table:

• Batch requests are 50% off input and output on every model.
• Cache reads cost 10% of base input on Haiku 5.5, 5% on Sonnet 5.5 and Opus 5.5, and 2.5% on Fable 5.1.
• Long context: Opus 5.5, Sonnet 5.5 and Fable 5.1 bill the full 1M window at standard per-token rates. Haiku 5.5 is the exception: a request whose prompt is over 100,000 tokens pays $0.50/$2.50, cache reads and writes included in that count.
• Tokenizer: all four models use the newer tokenizer Anthropic introduced with Claude Opus 4.7. If you are moving from Haiku 4.5, recount your prompts: Anthropic says the same text counts as about 30% more tokens on Haiku 5.5.
• Cloud endpoints: Bedrock and Google Cloud regional endpoints carry a 10% premium over global endpoints, per the same page.

A quick worked example at list price. A workload of 10 million input tokens and 2 million output tokens a month, in requests under 100K tokens each, costs $2 on Haiku 5.5, $40 on Sonnet 5.5, $80 on Opus 5.5 and $200 on Fable 5.1. Run it through the Batch API and each figure halves.

Opus deep dive

Claude Opus 5.5 is the model to use when a wrong answer costs more than the call. Anthropic built it "for long-running agentic coding and knowledge work" and priced it at $4 input and $20 output per million tokens, twice Sonnet 5.5. It has a 1M context window, 128K max output and moderate latency.

Anthropic's selection matrix maps Opus 5.5 to "multihour autonomous coding agents, large-scale refactoring, complex systems engineering, vision-heavy workflows, computer use." Those are the jobs where the price gap matters least, because one good plan or one caught bug is worth far more than the extra tokens.

Where Opus 5.5 fits in a routed stack:

• Deep refactors across many files. Large-scale refactoring is on Anthropic's own list for Opus 5.5.
• Plan-then-execute agents. Opus plans, Sonnet or Haiku executes. The planner fires a few times per task; the executors do the volume. We run this pattern in production.
• Code review. A review pass reads a lot and writes little, and Opus 5.5 cache reads cost $0.20 per million tokens against $0.10 on Sonnet 5.5. On cached context the bill is mostly reads, so the gap stays at about 2x on both input and output.
• Multihour agents. Opus 5.5 supports fast mode in research preview, which Anthropic says "delivers up to 2.5x higher output tokens per second from Claude Opus 5.5, Claude Opus 5, and Claude Opus 4.8 at premium pricing."

Two changes to check before you switch. On Opus 5.5, thinking can't be disabled and forced tool use (tool_choice of any or tool) returns an error, per Anthropic's model page. Sonnet 5.5 has the same forced-tool-use restriction. Many classifiers rely on forced tool use; Haiku 5.5's migration notes don't list that restriction, which is one more reason to keep triage on Haiku.

Where Opus is the wrong choice: high-volume traffic, sub-second responses, and any task where Sonnet 5.5 passes your evals at half the price.

Sonnet deep dive

Claude Sonnet 5.5 is the model you build your product on. Anthropic calls it "the best combination of speed and intelligence" and maps it to code generation, data analysis, content creation, visual understanding and agentic tool use. It costs $2 input and $10 output per million tokens, with a 1M context window, 128K max output and fast latency.

Sonnet 5.5 shipped on September 28, 2026 at the same list price as Sonnet 5, so the per-token bill doesn't change on fresh input and output. Cache reads drop from $0.20 to $0.10 per million tokens, half of Opus 5.5's $0.20, and the tokenizer is the same as Sonnet 5's, so your token counts carry over.

Three breaking changes to check before you switch from Sonnet 5, per Anthropic's what's-new page: thinking: {"type": "disabled"} returns a 400 error (the lowest setting is now between_tools), forced tool use (tool_choice of any or tool) returns a 400 error, and thinking blocks are tied to the model and conversation that produced them. For schema-valid tool calls, Anthropic points to strict: true tool use or structured outputs instead of forced tool use.

Where Sonnet 5.5 is the right default:

• Claude Code as the daily driver. This is the workflow our Claude Code agency practice deploys most often. Sonnet does the everyday work in the IDE and hard problems get escalated to Opus.
• Customer-facing chatbots. Fast latency and a $2/$10 price fit real product UX.
• Agent SDK builds. Agentic tool use is on Anthropic's list for Sonnet 5.5. Production agents we ship usually run Sonnet for the main loop.
• Document analysis and RAG. Data analysis is on the same list, and the 1M context window bills at standard rates.

Where Sonnet is wrong: multihour autonomous runs and large-scale refactors, which Anthropic maps to Opus 5.5, and narrow high-volume work where Haiku 5.5 costs a twentieth as much per token.

Haiku deep dive

Claude Haiku 5.5 is the triage model. At $0.10 input and $0.50 output per million tokens (for prompts up to 100,000 tokens) it is the cheapest current Claude model by a wide margin, and Anthropic labels its latency "Fastest." Anthropic builds it "for high-volume, latency-sensitive work such as classification, routing, extraction, and subagent tasks."

Haiku 5.5 was released on October 7, 2026. Compared with Haiku 4.5 it moves from a 200K to a 1M context window and from 64K to 128K max output, and it adds adaptive thinking with the effort parameter (default medium). The minimum cacheable prompt drops from 4,096 to 512 tokens, which helps short classifier prompts.

Anthropic's own "efficiency-first" advice starts here: begin with Haiku 5.5, test your use case, and "upgrade only if necessary for specific capability gaps." Anthropic says this approach fits prototyping, tight latency requirements, cost-sensitive work and "high-volume, straightforward tasks."

Where Haiku 5.5 fits:

• Triage and routing. First-line classification before handing off to a heavier model.
• Batch enrichment. Cleaning, labeling and normalizing thousands of records at $0.05/$0.25 in batch.
• Sub-agent workers. In a multi-agent system, the orchestrator runs Sonnet or Opus; the workers run Haiku.
• Tight tool calls. When the schema is fixed and the task is "call function X with these arguments."
• Real-time UX. Autocomplete-style features and instant suggestions, where latency matters more than depth.

Before you migrate from Haiku 4.5, check Anthropic's breaking changes: manual extended thinking (budget_tokens), non-default temperature, top_p or top_k, and assistant-message prefill all return errors. Recount your prompts too, because the newer tokenizer counts the same text as about 30% more tokens. Where Haiku breaks: long autonomous runs, deep multi-step reasoning, and prompts over 100,000 tokens, which move to the higher $0.50/$2.50 price.

Where Fable 5.1 fits

Claude Fable 5.1 is not part of the Haiku, Sonnet and Opus ladder most people search for, but Anthropic now lists it above Opus. Released September 1, 2026 at $10 input and $50 output per million tokens, it costs 2.5 times Opus 5.5 per token. Anthropic positions it as the step to take when Opus 5.5 still falls short.

Anthropic calls Fable 5.1 its "most capable model open to all customers" and recommends it "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." Its cache reads cost $0.25 per million tokens, 2.5% of base input. Claude Mythos 5.1 has the same specs and prices but is invitation-only through Project Glasswing.

In practice Fable 5.1 is an escalation path, not a fourth default tier. For a full cost breakdown, see Claude Fable 5 pricing explained.

The routing rule: triage, build, review

Route by job, not by habit. Haiku 5.5 triages because it is the cheapest and fastest model. Sonnet 5.5 builds because Anthropic maps everyday coding and tool use to it at half the Opus price. Opus 5.5 reviews and runs long autonomous jobs. Fable 5.1 is the exception path when Opus 5.5 at high effort still fails your evals.

Two caveats from Anthropic's own guide. First, "tuning effort is often a better lever than switching models." Opus 5.5 defaults to medium effort, and dropping effort can cut cost without changing models. Second, the guide says "most workloads start with Claude Opus 5.5." If you have no evals yet, start there, then move steps down to Sonnet 5.5 and Haiku 5.5 wherever your scores hold. Anthropic describes the same idea as combining models: an executor that escalates hard decisions to an advisor, or an orchestrator that delegates bulk work to lower-cost workers.

Head-to-head: which task each wins

This task map follows Anthropic's selection matrix and the verified prices. Sonnet 5.5 is the default for building, Opus 5.5 for large refactors, reviews and multihour agents, and Haiku 5.5 for classification, extraction and latency-bound work. Treat each pick as a starting point to test against your own evals, not a measured result.

Writing a new feature in a real codebase. Sonnet 5.5. Escalate to Opus 5.5 when the feature touches many files or has architectural implications.

Refactoring a complex module across many files. Opus 5.5. Large-scale refactoring is on Anthropic's list for it. Sonnet 5.5 covers everyday refactors.

Writing unit tests for an existing module. Sonnet 5.5. Haiku 5.5 can take simple, well-specified cases.

Code review on a non-trivial PR. Opus 5.5. Review is read-heavy, and cached reads cost $0.20, so long reviews stay affordable.

Classifying an incoming support ticket into one of 20 categories. Haiku 5.5. It is the cheapest and fastest option, and Opus 5.5 and Sonnet 5.5 reject forced tool use.

Long-running autonomous coding agent (30+ steps). Opus 5.5 plans, Sonnet 5.5 executes. Anthropic maps multihour autonomous coding to Opus 5.5.

Customer-facing chat assistant. Sonnet 5.5, for fast latency at $2/$10.

Extracting structured data from 50,000 PDFs. Haiku 5.5 through the Batch API, at $0.05/$0.25 for prompts up to 100K tokens.

RAG question-answering over a 200-document corpus. Sonnet 5.5 for most queries, Opus 5.5 for the hardest ones. If the retrieved context runs past 100K tokens, Haiku 5.5 moves to its higher $0.50/$2.50 price.

Voice assistant where latency matters more than depth. Haiku 5.5. Speed is the product.

Generating production-quality marketing copy. Sonnet 5.5 (content creation is on its list). Opus 5.5 for the hero pages where every word counts.

Building an internal tool with Claude Code over a week. Sonnet 5.5 for the daily work, Opus 5.5 on demand for the hard parts. This is the pattern most of our Claude Code agency engagements settle into.

Cost optimization patterns

The biggest avoidable cost is sending every request to one model. At list price, moving a workload from Opus 5.5 to Sonnet 5.5 halves the token bill, and moving it from Sonnet 5.5 to Haiku 5.5 cuts it by about 95%. The Batch API takes another 50% off, and cache reads cost 2.5% to 10% of base input.

Pattern 1: Triage, build, review. Haiku reads incoming work and decides what kind of task it is. Sonnet does the work. Opus reviews the output before it ships. This fits any high-stakes agent: coding assistants, customer ops automation, content pipelines. Haiku handles the volume at the lowest price, Sonnet does the bulk of the generation, and Opus only sees finished work.

Pattern 2: Planner and executor split. Opus generates the plan. Sonnet executes each step. Haiku handles narrow tool calls within each step. The planner runs once or twice per task; the executor runs many times, so most tokens bill at Sonnet or Haiku rates. We use this pattern for long-horizon coding agents.

Pattern 3: Confidence-gated escalation. Start every request on Haiku. If confidence falls below a threshold, or a downstream validator catches an error, escalate to Sonnet. If Sonnet is still unsure, escalate to Opus. This works well for classification and extraction at scale, where most requests are easy.

Pattern 4: Batch everything that can wait. Overnight enrichment, report generation and backfills do not need a real-time answer. Send them through the Batch API at 50% off: Opus 5.5 drops to $2/$10 and Haiku 5.5 to $0.05/$0.25.

Pattern 5: Prompt caching on top. Independent of model choice, cache hits bill at 10% of base input on Haiku 5.5, 5% on Sonnet 5.5 and Opus 5.5, and 2.5% on Fable 5.1. Cache writes cost more than base input (a 5-minute write on Sonnet 5.5 is $2.50 per million tokens), so caching pays off on prefixes you reuse.

The teams that hit Claude's pricing wall are usually running pattern zero: everything goes to one model. Routing is the fix, and your own evals tell you how far down the ladder each task can go.

Example routing architectures

These are example architectures with list-price math, not case studies. Each uses the triage, build, review split. Token counts are stated assumptions so you can swap in your own numbers; the prices are Anthropic's Claude API list rates as checked on October 23, 2026.

Customer support triage at a SaaS company. Inbound tickets hit Haiku 5.5 for category, urgency and sentiment. Routine tickets get a Haiku draft that a human reviews. Nuanced tickets go to Sonnet 5.5 for the reply. Tickets flagged as legal, contractual or escalation-risk go to a human queue with an Opus 5.5 brief attached.

Codebase migration agent. Opus 5.5 writes the migration plan once. Sonnet 5.5 does the per-file rewrites. Haiku 5.5 handles narrow jobs like "rename this import everywhere it appears" or "extract the function signatures from this file." The Opus share of tokens is small, so the bill tracks Sonnet and Haiku rates.

Sales enrichment pipeline. Assume 50,000 companies at about 1,000 input tokens and 100 output tokens each: 50 million input and 5 million output. On Haiku 5.5 through the Batch API that is $2.50 + $1.25 = $3.75. The same job costs $150 on Sonnet 5.5 and $300 on Opus 5.5 at standard rates. Send only the low-confidence edge cases up to Sonnet.

Multi-agent research assistant. An Opus 5.5 orchestrator decides which sub-agents to spawn. Each sub-agent runs Sonnet 5.5 and researches one angle. Sub-agents call narrow Haiku-powered tools to fetch pages, parse PDFs and extract tables. Planning bills at Opus rates; the long tail bills at Sonnet and Haiku rates.

Coding assistant inside an IDE. Sonnet 5.5 handles interactive completions and chat. A "deep refactor" or "review my PR" command routes to Opus 5.5. Haiku 5.5 powers fast inline suggestions that run constantly.

Pick the pattern that matches your workload. Default-routing is the trap.

Want help routing Claude in production?

We design and ship Claude-native systems for teams that need a routed, cost-optimized stack in production. That includes picking the right model per task, wiring prompt caching and batch jobs, building agent loops with the right orchestrator and executor split, and shipping it inside Claude Code workflows your team can maintain.

If your Claude bill is bigger than it should be, or an agent is slower than it should be, talk to our AI agent development and Claude Code agency team. We will scope the right routing pattern, ship it, and hand off the playbook. Book a consultation and we will look at your current setup.

FAQ

What is the actual difference between Opus, Sonnet, and Haiku?

They are three tiers of the same Claude family, tuned to different points on the price, speed and capability curve. As of October 23, 2026, Opus 5.5 costs $4/$20 per million input/output tokens, Sonnet 5.5 costs $2/$10 and Haiku 5.5 costs $0.10/$0.50 for prompts up to 100,000 tokens. All three have a 1M context window and 128K max output.

How much does Claude Opus 5.5 cost compared with Sonnet 5.5 and Haiku 5.5?

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, twice Sonnet 5.5 ($2/$10) and 40 times Haiku 5.5 ($0.10/$0.50 for prompts up to 100,000 tokens). In the Batch API the prices halve to $2/$10, $1/$5 and $0.05/$0.25. Cache reads cost $0.20 on Opus 5.5, $0.10 on Sonnet 5.5 and $0.01 on Haiku 5.5.

What is the difference between Haiku 4.5 and Haiku 5.5?

Haiku 5.5, released October 7, 2026, lists at $0.10/$0.50 per million tokens against Haiku 4.5's $1/$5, with a 1M context window instead of 200K and 128K max output instead of 64K. It adds adaptive thinking and the effort parameter. Two cost caveats: prompts over 100,000 tokens pay $0.50/$2.50, and the newer tokenizer counts the same text as about 30% more tokens.

Is Sonnet 5.5 more expensive than Sonnet 5?

No. Both list at $2 input and $10 output per million tokens, and they share a tokenizer. Sonnet 5.5 cuts cache reads from $0.20 to $0.10. Before switching, note that Sonnet 5.5 rejects forced tool use and thinking: {"type": "disabled"}.

Is Claude Opus 5.5 cheaper than Opus 4.8?

Yes, per token. Opus 5.5 lists at $4 input and $20 output per million tokens, while Opus 4.8 and Opus 5 list at $5 and $25. Cache reads drop from $0.50 to $0.20, and batch prices drop from $2.50/$12.50 to $2/$10. The old $15/$75 Opus price belongs to Opus 4 and 4.1, now retired on Anthropic's API.

Should I just always use Opus to be safe?

Not by default. Opus 5.5 costs twice Sonnet 5.5 per token, so running it on everything doubles the bill for work Sonnet may handle. Anthropic's guide does suggest starting on Opus 5.5 when you have no evals. Build a small eval set, then move each task to the cheapest model that passes it.

Should I use Fable 5.1 instead of Opus 5.5?

Only when Opus 5.5 falls short. Anthropic recommends Fable 5.1 for demanding reasoning and long-horizon agentic work, "or when your evals on Claude Opus 5.5 at higher effort still fall short." It costs $10/$50 per million tokens, 2.5 times Opus 5.5, with slower latency. Try raising Opus 5.5's effort level before you switch.

Is Haiku good enough for production?

For the right tasks, yes. Anthropic recommends Haiku 5.5 for real-time applications, high-volume processing, cost-sensitive deployments and sub-agent tasks. It fits classification, routing, batch enrichment and narrow tool calls. It is the wrong pick for long autonomous agent runs, and prompts over 100,000 tokens move to its higher price. Test it on your own data before deciding.

Which model should I use inside Claude Code?

Sonnet 5.5 is a sound default for everyday coding work inside Claude Code: Anthropic maps code generation and agentic tool use to it at $2/$10. Escalate to Opus 5.5 for large refactors, architecture decisions and PR review. Haiku 5.5 fits agent-side tooling and sub-tasks better than primary coding work.

How do I cut my Claude bill without losing quality?

Four levers. Route: send each task to the cheapest model that passes your evals. Cache: cache hits bill at 10% of base input on Haiku 5.5 and 5% on Sonnet 5.5 and Opus 5.5. Batch: the Batch API takes 50% off input and output. Tune effort: Anthropic says it is often a better lever than switching models.

Is Claude better than GPT for coding in 2026?

It depends on the task and on your own evals, and this guide only covers Claude's lineup. Sonnet 5.5 and Opus 5.5 are Anthropic's recommended models for coding and agentic work. Our Claude vs ChatGPT for coding comparison walks through the head to head in detail.

Do all three models support tool use and vision?

Yes. Anthropic's models overview says all current models support text and image input, text output, multilingual capabilities, vision and tool use. The differences are in thinking modes, latency, context and price. Opus 5.5 and Sonnet 5.5 do not accept forced tool use (tool_choice of any or tool), which matters for schema-bound classifiers.

Can I switch models without changing my code?

Mostly. The Anthropic API uses the same request shape across models, so you usually change a model string. Check the breaking changes first: on Opus 5.5 and Sonnet 5.5, forced tool use returns an error and thinking can't be fully disabled; on Haiku 5.5, manual thinking budgets, non-default sampling parameters and assistant prefill return errors. Prompts may need light tuning between tiers. Building a routed system that picks the right model per request is a separate architecture concern, and our AI agent development service and Claude Code agency practice do this routing design as a standard part of production builds; book a consultation if you want help.

Sources

Checked October 23, 2026:

• Anthropic, Pricing: model prices, cache reads, batch prices, long-context pricing (including Haiku 5.5's 100K threshold).
• Anthropic, Models overview: descriptions, latency labels, context windows, max output, retirement commitments.
• Anthropic, Choosing the right model: selection matrix, effort guidance, efficiency-first and capability-first approaches.
• Anthropic model pages for Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, plus the what's-new pages for Sonnet 5.5 and Haiku 5.5: release dates and breaking changes.