Which Claude model should you use? The current lineup and official prices

Anthropic shipped three new Claude models in 16 days: Claude Opus 5.5 on September 22, 2026, Claude Sonnet 5.5 on September 28, and Claude Haiku 5.5 on October 7. Claude Fable 5.1, released September 1, sits above all three. The result is a four-model lineup with a clear default and some pricing quirks you should know about before you switch a production workload. This guide lists what's current, what each model costs according to Anthropic's own pricing page, and how to choose. It also covers the API calls you need to check models and control spend. If you're weighing Claude against OpenAI's newest models, read our coverage of the GPT-6 launch alongside this.

Last verified: October 8, 2026 (prices and model status change often, so recheck the linked official pages before you commit).

The short answer

• Default to Claude Opus 5.5. Anthropic's models overview recommends it "for most workloads". It is also cheaper per token than the Opus it replaces.
• Move to Claude Fable 5.1 only for demanding reasoning and long-horizon agent work, or, in Anthropic's words, "when Opus 5.5 at higher effort still falls short on your evals". It costs 2.5 times as much as Opus 5.5.
• Use Claude Sonnet 5.5 when you need speed and lower cost but still want strong coding and agent performance. It costs half as much as Opus 5.5.
• Use Claude Haiku 5.5 for classification, extraction, routing, summaries and subagents. It is very cheap for prompts under 100,000 tokens.
• Claude Mythos 5.1 is not something you can simply choose. Access is restricted to Project Glasswing participants and, since October 6, 2026, organizations admitted to Anthropic's Cyber Verification Program.

The current lineup at a glance

All four current models accept text and image input, support tool use, have a 1M-token context window and 128K maximum output, and have a "reliable knowledge cutoff" of June 2026, according to Anthropic's models overview.

"Always on" matters. You can't turn thinking off on Opus 5.5 (the September 22 release notes list this as a breaking change), so you pay for some reasoning tokens even on simple requests. The effort setting is how you control that. More on this below.

Official API prices

Prices are in US dollars per million tokens (MTok), from Anthropic's pricing page. "Cache write" is the 5-minute cache; a 1-hour cache write costs 2x base input.

Last verified: October 8, 2026

What to notice:

• Opus got cheaper. Opus 5 was $5/$25; Opus 5.5 is $4/$20. Anthropic also claims Opus 5.5 costs "about 40% lower" than Opus 5 on typical workloads at default settings, because it uses fewer tokens per task. That is a vendor claim, so measure it on your own traffic.
• Sonnet 5.5's cache reads were halved on October 7, 2026, from $0.20 to $0.10. Anthropic's September 28 launch page still shows $0.20, but the pricing page and release notes show $0.10.
• Haiku 5.5 is priced by prompt length. Under 100K tokens it costs a tenth of Haiku 4.5 ($1/$5). Over 100K tokens the whole request moves to the higher tier, five times the price. Anthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average: about 90% less on short requests and about 50% less on long ones.
• Cache reads are discounted more than usual on the new models. The standard cache-hit rate is 0.1x base input. Opus 5.5 and Sonnet 5.5 charge 0.05x, and Fable 5.1 and Mythos 5.1 charge 0.025x.
• Batch and caching stack. The Batch API halves input and output prices, and the discount combines with prompt caching.
• Extras. Fast mode for Opus 5.5 (research preview, Claude API only) costs $8 input and $40 output, which is 2x standard; Anthropic says it is up to 2.5x faster. US-only inference (inference_geo: "us") multiplies all token prices by 1.1x. Web search costs $10 per 1,000 searches plus tokens.

The other cost you won't see in the table is the tokenizer. Anthropic's pricing page says Opus 4.7 and later models "produce approximately 30% more tokens for the same text", while per-token prices are unchanged. If you're moving from Sonnet 4.6 or Haiku 4.5, a lower per-token price doesn't mean a proportionally lower bill. We work through the arithmetic in why token pricing isn't the story.

How to choose: a decision guide by use case

Two cautions. First, Anthropic's benchmark numbers for these models are vendor-reported, so test on your own tasks. Second, mixing models in one product is a design decision of its own. Our explainer on model routing covers when sending easy requests to Haiku and hard ones to Opus actually saves money.

Step by step: check models and set effort via the API

The examples use the official Python SDK. We documented against anthropic 1.12.1 (released October 8, 2026, requires Python 3.10+).

1. Install and set your key.

pip install --upgrade anthropic
export ANTHROPIC_API_KEY="sk-ant-..."

2. List the models your key can use. The Models API returns each model's id, display_name, max_input_tokens, max_tokens and a capabilities object. By default it lists active and deprecated models.

curl https://api.anthropic.com/v1/models \
    -H 'anthropic-version: 2023-06-01' \
    -H "X-Api-Key: $ANTHROPIC_API_KEY"

Read max_input_tokens and capabilities from the API rather than hard-coding them. Anthropic adds fields regularly. In October 2026 it added capabilities.thinking.types.disabled and capabilities.server_tools.

3. Count tokens before you estimate costs. Token counting is free (it has its own rate limits) and uses the tokenizer of the model you name. That matters because of the ~30% tokenizer difference.

import anthropic

client = anthropic.Anthropic()
count = client.messages.count_tokens(
    model="claude-opus-5-5",
    system="You are a scientist",
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
print(count.input_tokens)

4. Set effort to control cost and latency. Effort applies to all output tokens, including thinking and tool calls. Anthropic calls it "a behavioral signal, not a strict token budget". The levels are low, medium, high, xhigh and max. Opus 5.5 and Haiku 5.5 default to medium, and most other models default to high.

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Review this migration plan for risks: ..."}],
    output_config={"effort": "medium"},
)
for block in response.content:
    if block.type == "text":
        print(block.text)

What Anthropic's docs say to expect: a normal Messages API response. On Opus 5.5 it may include thinking blocks before the text, because thinking can't be disabled. Check response.usage to see what you were billed for.

If you want a full agent loop rather than single calls, our guide to building your first agent with the Claude Agent SDK picks up from here.

Troubleshooting common migration errors

The new models changed several defaults. These issues come from Anthropic's release notes:

• 400 error with tool_choice: "any" or "tool". Forced tool use is rejected on Fable 5.1, Opus 5.5 and Sonnet 5.5. Use auto with strict tool use instead.
• 400 error when disabling thinking on Opus 5.5. You can't. Lower the effort instead. On Sonnet 5.5, the release notes say to use thinking: {"type": "between_tools"} at high effort or below rather than "disabled".
• 400 error with budget_tokens on Haiku 5.5. Manual thinking budgets were removed. Use effort.
• inference_geo returns 400. It is supported only on Claude 4.6 and later models.
• Bills higher than expected after migrating. Recount your prompts with the new model's tokenizer (step 3) and check usage for thinking tokens.

In the Claude apps

The app plans are priced separately from the API. According to claude.com/pricing:

• Free: no Opus and no Claude Code.
• Pro: $20/month, or $17/month billed annually. Adds Opus and Claude Code.
• Max: from $100/month, with 5x or 20x Pro usage.
• Team: standard seats are $25/month ($20 billed annually) and premium seats are $125/month ($100 billed annually).

Fable models are on all paid plans, but on Pro and standard seats they run on pay-as-you-go usage credits from the start. On Max and premium seats you can spend up to 50% of your weekly limit on Fable, per Anthropic's Help Center. (That is standard plan access. An earlier Fable 5 promotion ended on July 19, 2026, and Fable 5.1 was never part of it.) Since October 7, Max plans also include monthly API credits ($100 for Max 5x, $200 for Max 20x), and Team plans get up to $500/month pooled. Anthropic's pages don't say which app plans offer Haiku 5.5 in the model picker. For the app itself, see our Claude app explainer.

In Claude Code on the Anthropic API, the opus, sonnet and haiku aliases point to Opus 5.5, Sonnet 5.5 and Haiku 5.5. Other providers differ, according to Claude Code's model-configuration docs (checked October 9, 2026). On Amazon Bedrock and Google Cloud, opus is Opus 5.5 but sonnet and haiku still resolve to Sonnet 4.5 and Haiku 4.5. On Claude Platform on AWS they resolve to Opus 5.5, Sonnet 4.6 and Haiku 4.5. On Microsoft Foundry they resolve to Opus 4.6, Sonnet 4.5 and Haiku 4.5. Pin full model IDs if you depend on a specific model.

Cloud availability

Anthropic says Opus 5.5, Sonnet 5.5 and Haiku 5.5 are available on Amazon Web Services, Google Cloud and Microsoft Azure. The Haiku 5.5 release note lists Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. On Bedrock and Google Cloud the cloud provider sets the price. Anthropic's docs say regional endpoints there cost 10% more than global ones. Fast mode is available only on Anthropic's own API.

Older models: what's still available and what's retiring

Last verified: October 9, 2026 (Anthropic's deprecations page still lists Haiku 4.5 as active, with no deprecation notice)

If you still run Sonnet 4.5, you have until November 30, 2026. Haiku 4.5 can be retired no sooner than October 15, 2026. Note that "not sooner than" is not an announced retirement, but Haiku 5.5 is both cheaper and, per Anthropic, more capable, so there's little reason to wait. For background on Anthropic's model strategy, see Anthropic explained.

What's confirmed and what isn't

• Confirmed (Anthropic's docs): model IDs, prices, context windows, maximum output, default effort, retirement dates, and Mythos 5.1 access routes.
• Vendor claims: "75% cheaper than Haiku 4.5", "about 40% lower cost than Opus 5", Sonnet 5.5 "costs up to 30% less for most work", and all benchmark scores.
• Not stated: which consumer plans offer Haiku 5.5 in the app, and Bedrock/Vertex list prices (set by those providers).