Claude Code Models: Which One to Use for Every Task in 2026

Quick Answer

Claude Code ships five current models: Opus 5.5, Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5. Opus 5.5 (claude-opus-5-5) shipped on September 22, 2026 at $4 per million input tokens and $20 output, 20% below Opus 5, and Anthropic reports it ahead of Fable 5.1 on its three published coding benchmarks. Opus 4.6 and Sonnet 4.6, the pair this page used to cover, are now legacy models you can still pin by ID. Which model you get by default is set by your plan, not your taste. Max, Team Premium, Enterprise and pay-as-you-go API accounts start on Opus 5. Pro and Team Standard start on Sonnet 5. Opus 5 costs $5 per million input tokens and $25 per million output. Sonnet 5 costs $2 and $10 for the same 1M token context window. Anthropic's own guidance is to start on Opus 5 and move up to Fable 5.1 only when your evals at xhigh or max effort still fall short. Opus 5.5 is not a plan default in the model-configuration docs we read, so select it by ID with /model claude-opus-5-5. Switch any time with /model. All figures checked September 28, 2026.

Claude Code Model Aliases: default, best, opus, opusplan, sonnet, haiku

Claude Code almost never wants a raw model ID. It wants an alias, and the alias resolves differently depending on which provider you are billed through. This is the table people are actually searching for.

Default model by plan

The default moved recently. Anthropic's model-configuration docs note that before Claude Code v2.1.219, default resolved to Opus 4.8 on the Anthropic API, Max, Team Premium and Enterprise pay-as-you-go. If you pinned a model months ago to dodge that, the pin is probably now holding you on an older model than the one you would get for free.

Benchmark Comparison

The honest headline first. Anthropic has not published a SWE-bench Verified score for Opus 5, Sonnet 5 or Fable 5.1. Not on the model pages, not in the launch posts. The SWE-bench Verified leaderboard stopped taking most submissions on November 18, 2025, when it began requiring an arXiv preprint and an author affiliated with an academic institution or an established research lab. Any SWE-bench Verified percentage you see quoted for these models is unsourced.

Here is what Anthropic did publish for Opus 5 on July 24, 2026, all of it stated as comparisons rather than absolute scores:

• Frontier-Bench v0.1: surpasses all other models and more than doubles Opus 4.8's performance at a lower cost per task
• CursorBench 3.2: at max effort, within 0.5% of Fable 5's peak score at half the cost per task
• ARC-AGI 3: three times the next-best model's score
• OSWorld 2.0: surpasses Fable 5's best result at just over a third of the cost
• Zapier AutomationBench: around 1.5x the next-best model's pass rate at the same cost per task

The Sonnet 5 launch post published no per-benchmark numbers at all. It says Sonnet 5 lands "close to that of Opus 4.8, but at lower prices" and points at the system card for detail.

Claude Opus 5.5 coding benchmarks

Anthropic's September 22, 2026 launch post for Opus 5.5 is the first in this generation to publish absolute coding scores for every Claude model it compares. All Opus 5.5 results use adaptive thinking at max effort, and Terminal-Bench 4.0 is reported at xhigh.

At its default medium effort, Opus 5.5 still scores 52.5% on CursorBench against 51.8% for Fable 5.1 at max, according to the same post. Anthropic adds its own caveat: at these levels the benchmark margins are a less reliable guide than they look, and in its internal use the gap to Fable 5.1 is narrower than the table suggests. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks.

What the independent numbers say

Those index scores are Artificial Analysis measurements, read on September 28, 2026, with adaptive reasoning at max effort on the three reasoning models. The gap between Fable 5.1 and Opus 5 is two points. The gap between Opus 5 and Sonnet 5 is thirteen. That shape matters more for model choice than any single coding percentage: the top two are close enough that price decides, and the drop to Sonnet is real but costs a fifth as much.

SlopCodeBench: the benchmark that penalizes over-writing

Dex Horthy at HumanLayer ran Opus 5 against SlopCodeBench on July 27, 2026. SlopCodeBench does not stop after one task. It runs a sequence of checkpoints on the same codebase and scores whether the code stays clean. On a 3-problem, 17-checkpoint subset, Opus 5 passed 24% (4 of 17) against Opus 4.8's 6% (1 of 17), with Opus 4.6 scoring 17% in the original paper. The finding that travelled further: Opus 5 wrote roughly five times as many functions as the other two models over the same set of challenges.

That is the failure mode to plan around. The newer model does not write worse code so much as more of it, and an agent loop that reviews its own output will amplify that. One commenter on the thread described a single CREATE TABLE migration ballooning into a 200-line script after a few rounds of automated review.

Pricing Per Model

You reach Claude Code models through a subscription plan or an API key. Subscription costs are fixed monthly. API costs scale with tokens.

API pricing (per million tokens)

Batch API requests take 50% off input and output on every model. Cache reads are 10% of the base input price everywhere except Fable 5.1, where Anthropic cut them to 2.5%. That is why Fable 5.1 costs the same per token as Fable 5 but less to run in a long agent session: the resent prefix is the bulk of the bill. Opus 5.5 takes the same route with a cache read of $0.20, 60% below Opus 5, and Anthropic says a typical workload at default settings costs 40% less than on Opus 5. The launch also raised five-hour usage limits on Pro, Max and Team.

Subscription plans

Pro now includes Opus, which was not true of the tier a year ago. The practical constraint on Pro is not model access, it is the usage ceiling: Pro defaults to Sonnet 5, and Opus burns the allowance faster. As of September 28, 2026 Claude's pricing page lists both Max tiers under a single "from $100/month" figure, so read the plan selector for the current 20x price rather than trusting a number from a blog post.

Speed Comparison

In an agentic loop the model makes many tool calls per task, so output speed compounds. These are Artificial Analysis measurements read on September 28, 2026, with adaptive reasoning at max effort on the three reasoning models.

Read the time-to-first-token column carefully. On the reasoning models it includes thinking time, and max effort is the setting that makes thinking longest. Sonnet 5 at max effort takes longer to say its first word than Opus 5 does, because it thinks harder to compensate. Drop to /effort medium and that column collapses. Haiku 4.5 does not support the effort parameter at all, which is why its 0.73s stands out.

Opus 5 and Opus 4.8 have a fast mode, a research preview that runs the same weights at up to 2.5x output speed at premium pricing. Opus 5.5 adds one in Claude Code at $8 per million input tokens and $40 output. Artificial Analysis had no Opus 5.5 speed row when we checked; Anthropic's own figure is output more than 30% faster than Opus 5. Claude Code v2.1.271 extended it to Remote sessions, where the host setting or a typed /fast applies if your organization allows it.

When to Use Fable 5.1

Fable 5.1 shipped September 1, 2026 and is the most capable model Claude Code can run. It costs $10 per million input tokens and $50 per million output, double Opus 5 on both legs. Thinking is always on and cannot be turned off: the session toggle, alwaysThinkingEnabled, and MAX_THINKING_TOKENS=0 all have no effect on Fable models.

Anthropic's stated rule is to reach for it in two cases:

• Sessions longer than one sitting: agent runs lasting hours, multistep research, analysis carried through to a finished document or deck
• When Opus 5 has already failed: your evals at xhigh or max effort still fall short

Select it with /model fable or claude --model fable, which needs Claude Code v2.1.257 or later. The fable alias resolves to Fable 5.1 unless you set ANTHROPIC_DEFAULT_FABLE_MODEL, with one exception: in Claude apps gateway sessions both fable and best resolve to Fable 5.

When to Use Opus 5.5 or Opus 5

Opus 5.5 is the first model to try on long, sprawling jobs. Anthropic's examples are codebase-wide: an early tester audited and fixed a 200,000-line codebase in under three hours where Opus 5 took over 20 hours and 2.5x the tokens, and an internal C-to-Rust port of HAProxy finished in 9.5 hours against 12 for Fable 5.1 at 51% lower cost. It is also more resistant than Opus 5 to prompt injection. Run it with claude --model claude-opus-5-5.

Opus 5 released July 24, 2026 and is where Anthropic says most workloads should start. It is the default for every plan above Pro. Effort defaults to high, and unlike its predecessor Opus 4.8, thinking is on by default.

• Multihour autonomous coding agents where the loop has to keep its own plan straight without a human turn
• Large-scale refactoring across a dependency graph too wide to hold in a single file read
• Complex systems engineering: concurrency, state machines, migration paths where the first attempt has to be right
• Vision-heavy workflows and computer use, where Anthropic reports Opus 5 beating Fable 5 on OSWorld 2.0 at just over a third of the cost
• Planning inside a hybrid session, which is what the opusplan alias automates

The caveat practitioners keep raising is verbosity. On an August 26, 2026 HN thread about Claude's output length, one developer described trying to rein in "opus 5's extremely disorganized and verbose output" with output styles and not getting far. Claude Code's built-in Concise output style, available from v2.1.237, is the supported lever there.

When to Use Sonnet 5

Sonnet 5 released June 30, 2026 and replaced Sonnet 4.6 as the daily driver. It is a drop-in upgrade from 4.6 with three behavior changes: adaptive thinking is on by default, manual extended thinking now returns a 400, and setting temperature, top_p or top_k to a non-default value returns a 400. Same 1M context window as Opus 5. Two fifths of the price.

• Feature development: endpoints, components, schemas, utility functions
• Code review and refactoring inside a single module
• Writing tests: unit, integration, end to end
• Bug fixes with clear reproduction steps
• Iterative work you plan to review over several rounds anyway
• Execution inside opusplan, which is exactly the slot Anthropic assigns it

One number worth knowing before you migrate a prompt: Sonnet 5 uses an updated tokenizer, and Anthropic says the same input costs roughly 1.0 to 1.35 times as many tokens as on Sonnet 4.6 depending on content type. Re-baseline your token budgets rather than assuming the price drop is the whole story.

When to Use Haiku 4.5

Haiku 4.5 is the odd one out in the current lineup. It is the only model with a 200K context window rather than 1M, the only one capped at 64K output, the only one that still uses manual extended thinking rather than adaptive, and the only one that does not support the effort parameter. Its knowledge cutoff is February 2025.

• Generating boilerplate from templates
• Formatting and lint fixes
• Simple file operations
• Commit message generation
• Quick explanations of a snippet
• Sub-agent fan-out, via CLAUDE_CODE_SUBAGENT_MODEL

Haiku 4.5 is also what ANTHROPIC_SMALL_FAST_MODEL slots into: the background calls Claude Code makes for conversation titles and summaries. Those run constantly and do not need a frontier model. Anthropic's retirement commitment for Haiku 4.5 runs to October 15, 2026, the nearest date in the lineup, so expect a successor.

How to Switch Models in Claude Code

Five ways to set the model. The first four win in the order listed.

1. Mid-session: the /model command

# Open the picker:
/model

# Or switch and save as your default:
/model opus
/model sonnet
/model haiku
/model fable
/model opusplan

# Clear a saved override and revert to your account default:
/model default

Typing /model <name> directly behaves like pressing Enter in the picker, which saves the choice. To switch for this session only, open the picker with bare /model and press s on the model's row. That distinction is not obvious and it is the reason people find themselves permanently on a model they meant to try once.

2. At launch: the --model flag

claude --model opus
claude --model sonnet
claude --model haiku
claude --model fable
claude --model opusplan

# Pin an exact snapshot instead of an alias:
claude --model claude-opus-5

3. Environment variable

export ANTHROPIC_MODEL=claude-opus-5

# Or for one session:
ANTHROPIC_MODEL=opus claude

# New sessions only, and only when nothing else selects a model:
export ANTHROPIC_DEFAULT_MODEL=opus

4. Settings file

{
  "model": "opus",
  "effortLevel": "high"
}

5. Effort, the lever most people skip

Effort trades intelligence for latency and cost inside a single model. Anthropic's guidance now says tuning effort is often a better move than switching models. Set it with /effort, the --effort flag, CLAUDE_CODE_EFFORT_LEVEL, or the effortLevel setting.

Level support varies. Fable 5.1, Fable 5, Opus 5, Sonnet 5, Opus 4.8 and Opus 4.7 take all of low through max. Opus 4.6 and Sonnet 4.6 take low, medium, high and max, with no xhigh.

The Hybrid Workflow

The cheapest workable setup is not one model. It is a different model per phase, with the expensive tokens spent where the decisions are made rather than where the code is typed.

The opusplan alias

Claude Code automates exactly that split. The opusplan alias runs opus in plan mode for reasoning and architecture, then switches to sonnet for code generation and implementation. One command:

/model opusplan

# With the 1M context window during planning (v2.1.265+):
/model opusplan[1m]

On third-party providers, ANTHROPIC_DEFAULT_OPUS_MODEL controls the planning half and ANTHROPIC_DEFAULT_SONNET_MODEL controls the execution half, so you can point each phase at a different Bedrock or Vertex deployment.

A practitioner variant worth knowing: an HN thread from July 30, 2026 asking which model to plan and code with had the original poster running "Claude Opus for planning at high effort and Claude Sonnet 5 for coding at max effort." The effort levels are inverted from what the price suggests, and that is deliberate. The planner needs breadth, the implementer needs care.

Using Non-Anthropic Models with Claude Code

Claude Code speaks the Anthropic Messages API. Any endpoint that serves that format can drive it. Morph serves /v1/messages for every open-source chat model it hosts, so Claude Code runs on Morph with four environment variables and a Morph API key.

export ANTHROPIC_BASE_URL="https://api.morphllm.com"
export ANTHROPIC_AUTH_TOKEN="YOUR_MORPH_API_KEY"
export ANTHROPIC_MODEL="morph-kimik3"
export ANTHROPIC_SMALL_FAST_MODEL="morph-glm53flash"
claude

Or persist it in ~/.claude/settings.json under an env block. ANTHROPIC_MODEL remaps the sonnet and opus slots. ANTHROPIC_SMALL_FAST_MODEL remaps haiku, which is where the background title and summary calls go.

Read, Edit, Write, Bash and Grep run through the standard tool_use and tool_result loop. Streaming, system prompts and cache_control work as Anthropic documents them. Model reasoning arrives as thinking blocks. Full setup is in the Morph Claude Code guide.

Remapping individual alias slots

On third-party providers you do not have to replace every model at once. Claude Code exposes one environment variable per alias slot, so you can leave opus pointed at Anthropic and move only the cheap slots:

export ANTHROPIC_DEFAULT_FABLE_MODEL=...
export ANTHROPIC_DEFAULT_OPUS_MODEL=...
export ANTHROPIC_DEFAULT_SONNET_MODEL=...
export ANTHROPIC_DEFAULT_HAIKU_MODEL=...
export CLAUDE_CODE_SUBAGENT_MODEL=...

For local models such as Ollama or LM Studio, the open-source Claude Code Router proxies Claude Code at several providers and adds /model provider,model_name switching.

What Developers Actually Say

Anthropic publishes comparisons. Developers publish workflows. Four reports from the last seven weeks:

The pattern across all four is that model tier is no longer the main dial. Effort is. A developer moving from Opus 4.8 at xhigh to Opus 5 at medium got a faster, cheaper session out of a newer model at a lower setting, which is precisely the tradeoff Anthropic's own selection guide now recommends testing before you change models.

Frequently Asked Questions

Model IDs, prices, context windows, aliases and plan defaults on this page were checked against platform.claude.com, code.claude.com, support.claude.com, the anthropic.com Opus 5.5 launch post, claude.com/pricing and artificialanalysis.ai on September 28, 2026. Morph prices are read directly from the pricing source in this repository, so they cannot drift from the billing system.