Quick answer
- Claude Opus 5.5 price
- $4 in / $20 out per MTok
- Claude Sonnet 5.5 price
- $2 in / $10 out per MTok
- Claude Haiku 5.5 price
- $0.10 in / $0.50 out per MTok (prompts up to 100K tokens)
- Terminal-Bench 4.0
- Sonnet 5.5 70.6% · Opus 5.5 66.4% (xhigh) · Haiku 5.5 39.2%
- FrontierCode v1.1
- Opus 5.5 54.4% · Sonnet 5.5 52.1% (xhigh) · Haiku 5.5 46.4%
- Claude Code default
- Claude Opus 5.5 (since v2.1.280)
The Three Tiers, Side by Side
The current Claude lineup is a clean ladder: Claude Haiku 5.5 (October 7, 2026) at the bottom, Claude Sonnet 5.5 (September 28, 2026) as the everyday workhorse, Claude Opus 5.5 (September 22, 2026) as Anthropic's recommended default, and Claude Fable 5.1 above all of them as the escalation for the hardest reasoning. All four have a 1M-token window and 128K maximum output, so the choice is capability against price, not whether the repo fits. This page is about the three tiers you will actually route most work between:
| Claude Opus 5.5 | Claude Sonnet 5.5 | Claude Haiku 5.5 | |
|---|---|---|---|
| Price (per MTok) | $4 in / $20 out | $2 in / $10 out | $0.10 in / $0.50 out |
| Price note | $0.20/MTok cache reads; Fast mode costs $8/$40 | $0.10/MTok cache reads; no long-context premium up to 1M tokens | $0.01 cached input; prompts above 100K tokens cost $0.50/$2.50 |
| Released | September 22, 2026 | September 28, 2026 | October 7, 2026 |
| Context window | 1M tokens (128K max output) | 1M tokens (128K max output) | 1M tokens (128K max output) |
| Anthropic's pick for | Complex agentic coding and enterprise work; most workloads start here | Speed and capability for everyday coding, agent and enterprise workloads | Lowest latency and price; sub-agent tasks |
| Terminal-Bench 4.0 | 66.4% (xhigh) | 70.6% | 39.2% |
| FrontierCode v1.1 (main) | 54.4% | 52.1% (xhigh) | 46.4% |
| CursorBench 4.0 | 57.8% | 55.5% | Not in our verified set |
| GDPval-AA v2.1 | 1846 Elo | 1844 Elo | 1620 Elo |
| Claude Code access | Current recommended default; medium effort by default | /model sonnet (Claude Code v2.1.284+); medium effort by default in Claude Code | /model haiku (Claude Code v2.1.293+); medium effort by default |
What Each Tier Is Actually For
Claude Opus 5.5: the recommended starting point
Anthropic's recommended default for most workloads, optimized for long-running agentic coding and knowledge work. Anthropic's guidance is unusually direct here: most workloads start with Claude Opus 5.5, and Claude Code has used it as the default model on Pro, Max, Team, Enterprise and API plans since v2.1.280. Its lead over Sonnet 5.5 is real but narrow, 2.3 points on FrontierCode v1.1 and 2.3 on CursorBench 4.0, so pay for it where a wrong turn is expensive. Full breakdown on the Opus 5.5 page.
Strengths
- Anthropic's recommended starting point for most workloads
- 1M-token context and 128K maximum output at $4/$20 per MTok
- More than 30% faster output than Opus 5 in Anthropic's tests
Trade-offs
- Fable 5.1 remains the escalation path for the hardest reasoning tasks
- Adaptive thinking is always on and cannot be disabled
- Fast mode doubles token prices
Best for
- Default Claude model for production coding
- Long-running migrations and repository-wide audits
- Knowledge work that needs strong quality without Fable pricing
Claude Sonnet 5.5: everyday work at half the price
Anthropic's fast everyday model: Opus-class agentic coding results at half Opus 5.5's price. It keeps the $2/$10 price of Claude Sonnet 5, halves its cache reads from $0.20 to $0.10, and posts the best Terminal-Bench 4.0 score Anthropic has published. The catch is on the API side: thinking can't be switched off, and forced tool_choice and non-default sampling parameters return 400 errors, so check your pipelines before swapping the model ID. Full breakdown on the Sonnet 5.5 page.
Strengths
- Highest Terminal-Bench 4.0 score Anthropic has published: 70.6%, ahead of Opus 5.5's 66.4%
- Two Elo points behind Opus 5.5 on GDPval-AA v2.1 (1844 vs 1846) at $2/$10 per MTok
- More than 30% faster than Sonnet 5 and up to 30% cheaper per task (Anthropic)
Trade-offs
- Behind Opus 5.5 on FrontierCode v1.1 (52.1% at xhigh vs 54.4%) and CursorBench 4.0 (55.5% vs 57.8%)
- Thinking can't be switched off; between_tools is the lowest setting
- Rejects non-default temperature, top_p and top_k, and forced tool_choice
Best for
- Well-scoped features, bug fixes and refactors at volume
- Executing a plan an Opus 5.5 session wrote (the opusplan pattern)
- Terminal-heavy agent work, where it posts Anthropic's best Terminal-Bench score
Claude Haiku 5.5: subagents and mechanical work
Anthropic's fastest and cheapest model, the first Haiku with adaptive thinking, effort levels, and a 1M-token window. It is the newest model in this comparison (October 7, 2026), and the 1M-token window removes the old small-context penalty of the Haiku tier. For fan-out work where you spawn ten cheap subagents, nothing else in the family comes close on cost, and Anthropic itself pitches it as a subagent under Opus and Sonnet. At 39.2% on Terminal-Bench 4.0 against Sonnet 5.5's 70.6%, it is still not the model to lead complex autonomous feature work. Full breakdown on the Haiku 5.5 page.
Strengths
- Cheapest Claude by far: $0.10/$0.50 per MTok for prompts up to 100K tokens
- 1M-token context and 128K output, up from Haiku 4.5's 200K window
- Adjustable effort from low to max, a first for the Haiku line
Trade-offs
- Prompts above 100K tokens cost 5x more ($0.50/$2.50)
- Well behind Sonnet 5.5 on Terminal-Bench 4.0 (39.2% vs 70.6%)
- Thinking is on by default, so tiny max_tokens budgets written for Haiku 4.5 need raising
Best for
- Subagents under an Opus 5.5 or Sonnet 5.5 lead session
- Classification, extraction, routing, and summaries at volume
- Bounded mechanical edits with clear acceptance criteria
Your AI workspace for shipping software.
Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.
Download AIDEN freeFree to start · macOS 12+ · No credit card required
Cost per Feature: The Actual Numbers
Per-MTok prices are abstract; features are not. Take a realistic agentic feature, 300K input tokens of repo context, spec, and tool results, 60K output tokens of code and reasoning, and the tier choice turns into dollars:
| 300K in / 60K out | Claude Opus 5.5 | Claude Sonnet 5.5 | Claude Haiku 5.5 |
|---|---|---|---|
| Input cost | $1.20 | $0.60 | $0.03 |
| Output cost | $1.20 | $0.60 | $0.03 |
| Feature total | $2.40 | $1.20 | $0.06 |
| Total if requests exceed 100K-token prompts | $2.40 (no surcharge) | $1.20 (no surcharge) | $0.30 |
Claude Haiku 5.5's feature total uses its rate for prompts up to 100K tokens. A request above 100K tokens costs $0.50/$2.50 per MTok for the whole request, which is the last row: a subagent that reads a large slice of the repo in one prompt pays the higher rate. Opus 5.5 and Sonnet 5.5 have no long-context surcharge up to 1M tokens.
So: $2.40, $1.20, $0.06 per feature at list prices. Sonnet 5.5 costs half of Opus 5.5 on every rate (input, output and cache reads), so for the same token counts it is always half the bill. For scale, the same feature on Claude Fable 5.1 at $10/$50 per MTok runs about $6.00. The gaps look dramatic in percentage terms and small in absolute terms, which is exactly why the right question is not “which is cheapest” but “which tier fails least on this feature”: one failed $1.20 Sonnet run plus a $2.40 Opus retry costs $3.60, more than starting on Opus. The full ladder, including Fable 5.1, is mapped in the best Claude model for coding.
The Opusplan Pattern: Plan High, Implement Cheap
The pattern the community converged on is plan with Opus, implement with Sonnet, built into Claude Code's opusplan model setting, which uses Opus in plan mode and hands execution to Sonnet. The logic: architectural mistakes are the expensive kind, so spend the premium tokens where mistakes compound, then let the workhorse type out the plan. The 5.5 launch tables back the split: Opus 5.5 keeps its lead on FrontierCode and CursorBench, and the execution phase lands on Sonnet 5.5, the stronger Terminal-Bench model.
The math holds up. On our 300K-in/60K-out feature, suppose planning takes 50K input and 10K output tokens and implementation takes the remaining 250K/50K. All-Opus costs $2.40; Opus-plans-Sonnet-implements costs about $1.40, 42% less, while keeping the flagship on the decisions that matter. On a hard feature you flip it and stay on Opus throughout; on a routine one you skip Opus entirely and let Sonnet plan too.
Fleet Economics: Mixing Tiers Across Parallel Agents
The tier question changes shape once you run agents in parallel. A common setup is a Sonnet fleet with an Opus lead: the one genuinely hard feature goes to Opus, the four routine ones run on Sonnet simultaneously. Five features at our worked rates: $7.20 for the mixed fleet versus $12.00 for all-Opus, 40% cheaper for the same wall-clock time, because the features that did not need the flagship did not pay for it.
The catch is that per-feature routing is tedious to do by hand across terminal tabs. That is the part AIDEN automates: it runs your existing Claude Code CLI on a kanban board where each feature is a card with its own git branch and its own model assignment, so Opus takes the migration feature while Sonnet clears the backlog next to it. The full lineup, including the GPT-6 side, lives at AI models for coding.