- Home
- Learn
- Deep Dives
- Claude Sonnet vs Opus vs Haiku: Which Model to Use
Claude Sonnet vs Opus vs Haiku: Which Model to Use
Claude Sonnet vs Opus vs Haiku as a routing rule: which model for coding, writing and bulk jobs, with prices, latency and cost per job. Checked 16-11-2026.
Published 16-11-2026. Last verified on 16-11-2026 against Anthropic's model overview, pricing page and the launch posts for Opus 5.5, Sonnet 5.5 and Haiku 5.5. We re-check this guide monthly, because a new release can flip a rule.
Quick answer: Claude Sonnet vs Opus vs Haiku, which should you use?
This guide is for developers and team leads who run Claude for several kinds of work, coding, writing and bulk processing, and want a rule per job instead of one model for everything. Our rule, based on Anthropic's published numbers on 16-11-2026: run Claude Sonnet 5.5 by default, move to Opus 5.5 only for long unattended agent runs and the hardest reasoning and send short, repetitive jobs to Haiku 5.5. The reason is price against a small quality gap. Opus 5.5 costs $4 input and $20 output per million tokens, Sonnet 5.5 $2 and $10 and Haiku 5.5 $0.10 and $0.50 for prompts up to 100,000 tokens. On most benchmarks Anthropic publishes for both, Sonnet 5.5 sits within about three points of Opus 5.5 at half the price.
How we built this routing rule
Our method in three lines: prices and latency labels come from Anthropic's docs on 16-11-2026; quality differences come from Anthropic's own launch posts and are labelled vendor-reported; and every cost per job is our calculation on list prices, with the assumptions written out.
Our gate: a recommendation for the more expensive model needs a published difference on a benchmark that matches the job. Where two models are within three points of each other on that benchmark, we recommend the cheaper one. The briefing for this page asked us to run ten identical tasks on all three models. We had no model access in this production run, so those runs are not tested by us; the rule rests on vendor numbers until we can add our own.
The gate removed two things that most comparisons include. "Opus for writing" is out: no Anthropic benchmark isolates writing quality and on general knowledge work the two are close, so writing goes to Sonnet. And we left OSWorld 2.1 out of the comparison, because Anthropic's Sonnet 5.5 and Haiku 5.5 posts report Sonnet on different subsets (80.1% partial and 83.9% offline), which makes a fair three-way line impossible.
The three models side by side
All three current models have a 1M-token context window and 128K maximum output, according to Anthropic's models overview. They differ in price, speed and the depth of reasoning per answer. Benchmarks below are vendor-reported, from the Opus 5.5, Sonnet 5.5 and Haiku 5.5 launch posts:
| Opus 5.5 | Sonnet 5.5 | Haiku 5.5 | |
|---|---|---|---|
| Released | 22-09-2026 | 28-09-2026 | 07-10-2026 |
| Input and output per million tokens | $4 and $20 | $2 and $10 | $0.10 and $0.50 (prompts up to 100K) |
| Latency (Anthropic's label) | Moderate | Fast | Fastest |
| Default effort on the API | Medium, thinking always on | High | Medium |
| FrontierCode 1.1 (Main) | 54.4% | 52.1% (xhigh) | 46.4% |
| Terminal-Bench 4.0 | 66.4% | 70.6% | 39.2% |
| Humanity's Last Exam, with tools | 67.7% | 64.5% | 57.4% |
| Chartography, no tools | 64.4% | 61.6% | 46.4% |
Two things in this table shaped our rule. Sonnet 5.5 is within 2.3 to 3.2 points of Opus 5.5 on three of the four rows and ahead of it on Terminal-Bench 4.0, as Anthropic reports it. Haiku 5.5 trails by 6 to 31 points, which is a real gap for complex work but irrelevant for a job like tagging a support ticket. Haiku 5.5 prompts over 100,000 tokens cost $0.50 and $2.50, five times more, per the pricing page.
Vendor tables are a starting point. For independent numbers inside the session where you choose, your agent can pull model details and cost per task from Artificial Analysis through Agensi Tools, from $0.013 per call for model details and from $0.13 per call for model cost per task, or the text and coding leaderboards from LMArena, from $0.013 per call (starting prices read on 16-11-2026). You need no account or API key at either provider and a failed call is free.
Claude Sonnet vs Opus: when is Opus worth double?
Opus is worth double when the job is long, unattended and expensive to get wrong. For everything else, Sonnet wins on our gate, because the published gap is about three points at twice the price.
Where we would pay for Opus 5.5:
- Long agent runs without a human in the loop. First because errors compound: a small drop in per-step accuracy over a hundred steps becomes a failed run. Anthropic describes Opus 5.5 as the model "for long-running agentic coding and knowledge work".
- Multi-file changes in a large codebase. Second: one wrong assumption early in a migration costs hours of review.
- Hard analysis over a large document set. Third: Opus leads Sonnet by 3.2 points on Humanity's Last Exam with tools, the widest of the four gaps above.
Where Sonnet 5.5 is the better choice: interactive coding where you review each step, first drafts, summaries and most skill runs. The Opus effort default is medium with thinking always on; Sonnet defaults to high on the API. In our view that is one reason the two land so close on Anthropic's tables.
The mistake we would most want teams to avoid is one model for everything. If you design agents that hand work between models, the agents and orchestration collection has skills for that.
Claude Haiku vs Sonnet: when is the small model enough?
Haiku 5.5 is enough when the task has a clear right answer, short input and runs at volume: classification, extraction, routing and tagging. That is exactly the use Anthropic names for it in the models overview. At $0.10 and $0.50 per million tokens it costs a twentieth of Sonnet 5.5 for prompts up to 100,000 tokens.
Haiku is not enough when the task needs several reasoning steps. Anthropic's own table shows Haiku 5.5 at 39.2% on Terminal-Bench 4.0 against 70.6% for Sonnet 5.5, a gap of more than 31 points on agentic terminal work. For a coding agent that is disqualifying in our view.
Two practical limits we would keep in mind. First, the price jumps fivefold above 100,000 input tokens, so a Haiku job that reads whole documents can cost more than you expect. In our reading of the pricing page, cached input also counts toward that prompt length; check this for your own setup. Second, Anthropic says Haiku 5.5 is its first Haiku with an adjustable effort setting, so test the job at low effort before you scale it up. Customer-facing replies are a good Haiku job in our view: Email Reply Generator, $1.99, drafts replies by intent, tone and language, and Customer Review Insight Analyzer, $1.99, turns reviews into a report with themes and sentiment.
Which Claude model is best for coding?
Sonnet 5.5 for interactive coding, Opus 5.5 for long autonomous changes. On FrontierCode 1.1 Anthropic reports 54.4% for Opus 5.5 and 52.1% for Sonnet 5.5 at xhigh effort, a 2.3-point gap. On Terminal-Bench 4.0 Sonnet 5.5 is ahead. That puts Sonnet inside our three-point rule for coding, so it gets the default.
We would still put code review on Opus when the change is large or security-sensitive, because a missed issue in review is the most expensive error in the chain. AI Code Reviewer and Security Analyzer, free, runs an automated review and security pass on changes; run it on Sonnet for routine pull requests and on Opus for releases. The code quality and review collection has more, and our guide to the best Claude model for coding goes deeper.
Which Claude model is best for writing?
Sonnet 5.5. No Anthropic benchmark measures writing quality directly, and on the closest proxy, general knowledge work, Sonnet sits 3.2 points behind Opus on Humanity's Last Exam with tools and 2 Elo points behind on GDPval-AA v2.1 (1844 against 1846, per the Sonnet 5.5 post). Our gate therefore sends writing to the cheaper model.
What changes writing quality more than the model, in our view, is the process around it: a clear brief, a source list and an edit pass. Humanize Writing, $7, rewrites stiff AI text into natural language while keeping the facts, which works well as a Sonnet second pass. We would use Haiku only for short, templated copy such as product tags or subject lines. The writing and content collection has skills for each step.
Which model for bulk and batch jobs?
Haiku 5.5 first, Sonnet 5.5 if Haiku fails a quality check. For bulk work the price per job is multiplied by thousands, so the cheapest model that passes a spot check wins. Batch processing halves the price again: Sonnet 5.5 costs $1 input and $5 output per million tokens in batch and Haiku 5.5 $0.05 and $0.25 for prompts up to 100,000 tokens, per the pricing page on 16-11-2026.
Our process for a new bulk job: run a sample of 50 items on Haiku, check them by hand and only move up to Sonnet if the error rate is too high for the job. For data exports, Csv Report Exception Auditor, $12.99, checks CSV and Excel exports for anomalies and drafts follow-up emails; it is the kind of structured task we would try on Haiku first. More in the data and databases collection.
What does each job cost on each model?
Our calculation on list prices, standard (not batch), without caching, counted in each model's own tokens:
| Job | Opus 5.5 | Sonnet 5.5 | Haiku 5.5 |
|---|---|---|---|
| One coding task (40,000 in, 4,000 out) | $0.24 | $0.12 | $0.006 |
| One long-form draft (8,000 in, 3,000 out) | $0.092 | $0.046 | $0.0023 |
| 10,000 support tickets (1,000 in, 50 out each) | $50.00 | $25.00 | $1.25 |
The coding row is where the choice between Opus and Sonnet matters per task; the ticket row is where Haiku saves real money. Prompt caching changes all of these: Opus 5.5 and Sonnet 5.5 read cached input at 0.05 times the base price. To see your actual numbers, Agent Cost Scan, free, reads your logs and reports cost per run and the most expensive model. Our guide to reducing Claude Code token usage and our comparison of AI coding agent costs cover the wider picture, and the Claude API pricing guide has the full price list.
How do you switch models mid-session?
In Claude Code, type /model and pick another model; the change applies to the rest of the session. The --model flag sets the model for one session from the start and the model key in settings.json sets your default. On the API you change the model value per request, which is how most routing setups work: each call goes to the model that fits it.
Our pattern for a working day: start on Sonnet, switch to Opus for the one hard problem, switch back when it is solved. Leaving a session on Opus after the hard part is, in our view, the easiest way for costs to creep up unnoticed. Teams can limit which models people pick with the availableModels setting in a shared settings file.
How do you check this rule on your own work?
Run a small comparison before you commit a team to a model. Vendor benchmarks measure Anthropic's tasks rather than yours. A three-point gap on a public benchmark can be zero or ten points on your own work. Our four-step check takes an afternoon:
- Pick ten real tasks. First because everything else depends on it: take them from last month's work, not from a tutorial, and include two that went wrong before.
- Write a pass or fail rule per task. Second: decide before you run anything what a correct result looks like, for example "tests pass" or "every figure matches the source". Without that rule, people tend to prefer the longest answer.
- Run each task on Sonnet 5.5 and on one neighbour. Third: Opus 5.5 if you suspect Sonnet is too weak, Haiku 5.5 if you suspect it is overkill. Record pass or fail, seconds to finish and the cost from the usage data.
- Apply the three-point rule to your own numbers. Last: if the cheaper model passes as many tasks, or one fewer out of ten, use it.
Ten tasks is a small sample, so treat the result as a direction, not a proof, and repeat it after each new model release. This is the test we plan to run ourselves for the next version of this page.
Which model should each skill type run on?
Our routing table for skills, in the order of how often a typical team runs them:
- Writing, summaries and office tasks: Sonnet 5.5. First because they are the bulk of daily skill runs and sit inside the three-point rule.
- Coding skills with review by a person: Sonnet 5.5. Second for the same reason, with Opus for large or security-sensitive changes.
- Classification, extraction and templated replies: Haiku 5.5. Third: high volume, clear answers, a twentieth of the Sonnet price.
- Long autonomous agents: Opus 5.5. Last in frequency but first in cost of failure.
To automate this, Route Each Task to the Right Model, $12, sends easy tasks to a cheaper model and escalates only when a check fails. To replace our vendor-based rule with your own measurements, Agent Evals with Real Statistics, $12, sets up golden datasets, confidence intervals and paired comparisons. The LLM ops collection and the workflow automation collection have more skills for routing and pipelines.
We sell skills and pay-per-call tools that your agent uses through one MCP connection, and we name ours above. No vendor paid to be listed.
Your next step
Set Sonnet 5.5 as your default, write down the two or three jobs that go to Opus and move one bulk job to Haiku with a 50-item spot check. To pull model data into the same session where you decide, connect your agent once through Agensi MCP.
Written by Samuel Rose, Founder of Agensi. Next scheduled review: 16-12-2026.