Best Claude model for coding: start with Opus 5.5

Anthropic's model-selection guide doesn't crown a single winner. It pairs each model with a kind of work, and states plainly that most workloads start with Claude Opus 5.5. That makes it the most useful starting point available: it comes from the vendor, not from a synthetic leaderboard. The right-hand column is our reading of what each row looks like inside a Claude Code session.

All four current models share a 1M-token context window and 128K maximum output, use adaptive thinking, and accept effort from low to max, so capability, speed and price are what separate them. Prices are standard API rates per million tokens; up to 1M tokens there is no long-context surcharge, except on Haiku 5.5, which bills prompts above 100K tokens at a higher tier.

What the 5.5 generation changed

Between September 22 and October 7, Anthropic replaced the Opus, Sonnet and Haiku models in turn. The pattern is the same each time: equal or lower prices, much higher agentic-coding scores.

Inside Claude Code the switch is automatic once you update. If you call the API directly, the 5.5 models bring breaking changes worth checking before you swap the ID: thinking can't be turned off on Opus 5.5 (Sonnet 5.5's lowest setting is between_tools), non-default temperature, top_p and top_k return a 400, and Opus 5.5 and Sonnet 5.5 reject forced tool_choice. Details are on the Sonnet 5.5 and Haiku 5.5 pages.

If you never set a model, Claude Code uses its default, Claude Opus 5.5 on every Anthropic plan and on the API, unless your organization set its own. To choose deliberately there are three levers, from the most temporary to the most permanent:

1. 1

   For one session: /model

   /model opens the picker: Enter saves your choice as the default, s applies it to this session only. /model sonnet switches directly.
2. 2

   For one run: --model

   Pass an alias or a full model ID when you launch Claude Code. It works for non-interactive claude -p runs too.

   claude --model claude-opus-5-5
   claude -p --model haiku "Summarize the failing tests in this log"

   Copy
3. 3

   As your default: settings.json

   Set the model in ~/.claude/settings.json, or in .claude/settings.json to share it with a repository. A per-model effort level can live next to it.

   {
     "model": "opus",
     "modelSettings": {
       "claude-opus-5-5": { "effortLevel": "high" }
     }
   }

   Copy

opusplan and sub-agents: two models in one session

The opusplan alias uses Opus in plan mode and Sonnet when it executes, today Claude Opus 5.5 and Claude Sonnet 5.5. Sub-agents can run on their own model and effort through their frontmatter, or all of them at once through CLAUDE_CODE_SUBAGENT_MODEL:

# .claude/agents/searcher.md
---
name: searcher
description: Finds every call site of a function and summarizes how it is used
model: haiku
effort: low
---
Find every call site of the function you are given and summarize how it is used.

claude --model opusplan
export CLAUDE_CODE_SUBAGENT_MODEL=haiku

To change what an alias points at, for example on Bedrock or Google Cloud, set ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL or ANTHROPIC_DEFAULT_FABLE_MODEL. Admins can restrict the picker with availableModels in managed settings; Haiku models always stay available.

Model choice is half the tuning. Effort is the other half: more effort buys more planning and checking per step, and costs time and tokens. Anthropic's advice is to start at the default and adjust from your own evals, and it notes that changing effort is often a better lever than changing models.

/effort high                       # this model, saved as your default
claude --effort xhigh               # one run
export CLAUDE_CODE_EFFORT_LEVEL=high

ultrathink, max and Ultracode are different tools

Writing ultrathink in a prompt asks for deeper reasoning on that one turn without changing your saved level. max raises the level for every turn until you change it. Ultracode is a separate toggle in /effort with which Claude orchestrates dynamic workflows for substantive tasks, at any effort level. Admins can cap effort with maxEffortLevel. Effort levels also don't map one-to-one across generations: Sonnet 5.5's were recalibrated, so re-run a familiar task before carrying an old setting across.

Per-token prices only mean something once a task sits behind them. The calculator applies Anthropic's list prices to one agent task, including how Claude Code's prompt cache bills: the reused context as cache reads, the new context as cache writes. The presets are illustrative sizes for a multi-turn run, not measured averages, so put your own numbers in.

Two things the per-token view hides. First, cache pricing is not uniform: cache reads cost 2.5% of input on Fable 5.1, 5% on Opus 5.5 and Sonnet 5.5, and 10% on Haiku 5.5, so in long cached sessions Fable 5.1 costs less than its 2.5x list price suggests. Second, a long Claude Code session passes 100K tokens of context quickly, so Haiku 5.5 as the main model often pays its higher tier; as a sub-agent with a fresh context it usually stays under. And token price is not task price: a model that lands a change in one attempt beats a cheaper one that needs three.

Every rule on this page ends in the same place: the right model is a property of the task, not of the repository. A rename and a data-layer migration in the same codebase want different models and different effort, and a single default set once per project quietly overpays on one while underpowering the other.

There are three ways to act on that, from lightest to heaviest:

The same logic crosses vendors. Some tasks suit a GPT-6 model better than any Claude tier; the Codex side of this decision is in the best Codex model, and mixing both on one codebase is covered in model routing for coding agents.

Claude Code's aliases follow the newest model automatically, but full model IDs pinned in settings, environment variables, sub-agent frontmatter or scripts stay exactly where you left them. These are the dates on Anthropic's deprecation page (Bedrock and Vertex set their own):

Anthropic's own replacement for Sonnet 4.5 is Sonnet 5.5. For the other rows we list the current model of the same tier, which is the natural move for coding work. A quick search finds the local pins:

# user and project Claude Code config, plus agents and scripts in this repo
grep -rnE 'claude-(sonnet-4-5|haiku-4-5|opus-4-[15])|claude-(opus|sonnet)-4-2025' ~/.claude .claude . --exclude-dir=node_modules --include='*.json' --include='*.md' --include='*.sh' 2>/dev/null
env | grep -E 'ANTHROPIC_(MODEL|DEFAULT_[A-Z]+_MODEL)|CLAUDE_CODE_SUBAGENT_MODEL'

Anthropic published coding results with each 5.5 launch. These are vendor numbers from three different launch tables, so treat small gaps as noise; GPT-6 Astra's figures are as reported by OpenAI in Anthropic's Opus 5.5 table.

Two readings, both hedged. Inside Anthropic's lineup, Sonnet 5.5 is the surprise: it beats Opus 5.5 on Terminal-Bench 4.0 at half the price, while Opus 5.5 keeps the lead on FrontierCode and CursorBench, which is consistent with Anthropic keeping Opus as the default for complex work. Fable 5.1 does not top these coding tables; Anthropic positions it for long-horizon work where Opus runs out of road, which short benchmarks don't capture. Across vendors, Opus 5.5 posts higher scores than GPT-6 Astra at $4 / $20 against $10 / $50; the longer comparison is in Claude vs GPT for coding. Harnesses and effort settings differ between tables, so treat them as a prompt to test on your own repository.

Sources

• Anthropic: Choosing the right model
• Anthropic: Models overview
• Anthropic: Pricing
• Anthropic: Model deprecations
• Anthropic: Claude Opus 5.5 launch table
• Anthropic: Claude Sonnet 5.5 launch table
• Anthropic: Claude Haiku 5.5 launch table
• Claude Code docs: Model configuration
• Claude Code changelog