Claude Opus 5.5 review roundup: the daily driver for coding, what it costs and what you get
Claude Opus 5.5 launched on September 22, 2026 and is the model reviewers call their daily driver. System-card scores, Artificial Analysis's effort ladder, DataCamp's tests and Bilibili creators: where the money goes, and when it pays.
In short
- Claude Opus 5.5 is Anthropic's main flagship. List price is $4 in / $20 out per 1M tokens, with no surcharge across its 1M-token context.
- It is first on Bito's coding benchmark with 85.0%, above the dearer Fable 5.2. On the Artificial Analysis index it scores 58 at max effort and 51 at its default.
- The cost is per task: the same task often costs three to four times what it does on GPT-6.1 Sol. The value is the ceiling: scores GPT-6.1 Sol cannot reach at any price.
- Since Fable 5.2 launched, reviewers still recommend Opus 5.5 for daily use. On MoeRouter it is in the default, stable and ccmax-cheap groups, and it is the default choice for Claude Code.
On this page
What Claude Opus 5.5 is#
Claude Opus 5.5 was released on September 22, 2026, the same day as GPT-6 Sol, as the first model of the Claude 5.5 line. Anthropic pitches it as the lead model for long-running agentic coding and knowledge work: Fable 5.1's results on most work at Opus prices[1]. The Bilibili creator GenJi's video that night described Anthropic and OpenAI as making their big moves one right after the other[6].
| Claude Opus 5.5 | |
|---|---|
| List price per 1M tokens | $4 input, $20 output |
| Caching | Read $0.20; write $5 (5 minutes) or $8 (1 hour) |
| Fast mode | $8 / $40, up to 2.5x faster |
| Context and output | 1M-token context billed at standard rates throughout, with no long-context surcharge; up to 128K output tokens |
| Reasoning effort | Five levels, and thinking cannot be switched off; the default is medium |
| Model name on MoeRouter | claude-opus-5-5 |
| Endpoints on MoeRouter | Anthropic Messages; the ccmax-cheap group also offers Chat Completions |
Specifications and prices are as DataCamp compiled them; the default effort is as CodingFleet reports it[1][2].
What it costs on MoeRouter#
| Model | Group | Input | Output | Cache read |
|---|---|---|---|---|
claude-opus-5-5 | ccmax-cheap | CN¥4.8≈ US$0.72 | CN¥24≈ US$3.6 | CN¥0.24≈ US$0.036 |
default | CN¥5.2≈ US$0.78 | CN¥26≈ US$3.9 | CN¥0.26≈ US$0.039 | |
stable | CN¥5.2≈ US$0.78 | CN¥26≈ US$3.9 | CN¥0.26≈ US$0.039 |
Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. The same model takes other routes in another group and costs differently there; a key's group is chosen when the key is created. About groups Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().
The three groups differ. The default group picks the cheapest healthy route of several, so its price is not fixed; the stable group has one fixed route and a fixed price per model; ccmax-cheap is another set of routes. Look at the table before choosing; details in about groups.
Benchmarks#
Reported by Anthropic#
| Benchmark | Opus 5.5 | For comparison |
|---|---|---|
| SWE-bench Pro | 89.9 | Fable 5.2 93.0, Fable 5.1 81.2 |
| Terminal-Bench 4.0 (xhigh) | 66.4% | GPT-6 Astra 57.9%, Fable 5.2 67.6% |
| FrontierCode 2.0 (max) | 45.5 | Fable 5.2 49.1, GPT-6 Astra 42.8 |
| OSWorld 2.1 (offline set) | 87.2 | Fable 5.2 89.3, GPT-6 Astra 79.5 |
| GDPval-AA v2.1 (Elo) | 1866 | Fable 5.2 1831, GPT-6 Astra 1584 |
These are the Opus 5.5 figures given in the Fable 5.2 system card[3]. Two things stand out: on GDPval-AA, a knowledge-work evaluation, Opus 5.5 is above the dearer Fable 5.2; and on Terminal-Bench 4.0 the two are 1.2 points apart, which the author who compiled the table calls noise[3]. DataCamp adds that Anthropic ran with production safeguards on, so cybersecurity and biology tasks were rerouted and the scores are probably on the low side[1].
From third parties#
- Coding. On Bito's coding benchmark at high effort Opus 5.5 scores 85.0% and is first; Fable 5.2 scores 83.1%, GPT-6.2 Astra 77.5% and GPT-6.2 Sol 76.5%[4].
- General intelligence by effort level. On the Artificial Analysis Intelligence Index v4.3.2, Opus 5.5 scores 42, 51, 54, 56 and 58 from low to max, at $0.55, $1.34, $1.82, $3.46 and $5.98 per task. GPT-6.1 Sol tops out at 52 ($0.72). Opus 5.5 wins 13 of the 16 component evaluations, by the widest margins on knowledge-work documents, scientific coding and expert reasoning[2].
- Coding in a terminal. In the same data, Opus 5.5 scores 59.6% on Terminal-Bench 4.0 and GPT-6.1 Sol 56.1%[2].
Hands-on: where the money goes#
DataCamp, the day after launch, had Opus 5.5 and GPT-6 Sol each build a single-file Tetris in which gravity flips every 20 seconds. Both got the flip right. Opus 5.5 took 3 turns, 2 tool calls and $0.87; Sol took 6 turns, 9 tool calls and $0.26, and won narrowly on polish. The author notes that this is one run of one task[1].
DataCamp, a week later, compared it with GPT-6.1 Sol on a clinic scheduler: conflict logic was perfect on both, and it cost $1.11 on Opus 5.5 against $0.27 on Sol, with Opus 5.5 using more than twice the output tokens, mostly on thinking[5].
ComputingForGeeks found Opus 5.5 level with Fable 5.2 on correctness across three infrastructure tasks, at $0.208 and 86.5 seconds in total at medium effort against Fable 5.2's $0.446 and 113.5 seconds. The flaw it found: both OpenTofu files Opus 5.5 produced failed the format check[3].
程序员阿江-Relakkes (Bilibili) tested three models on launch day with a multi-robot scheduling simulation. His description says Opus 5.5 did the multi-robot task beautifully and GPT-6 Sol impressed him on speed and quota, and that this is not a standardised benchmark[7]. Six weeks later Fable 5.2 passed Opus 5.5 on his private question bank, and that video still ends on why he recommends Opus 5.5 for daily use[9].
实验式 (Bilibili) compared Gemini 4 Argon on 3D tasks and concluded that Argon still trails Claude Opus 5.5 overall, though it is pleasant to use, steady and fast[8].
Taken together: what makes Opus 5.5 expensive is not only the token price, twice GPT-6.1 Sol's, but the tokens it spends per task. The same small task often costs three to four times as much as on Sol.
What you get for it#
- The ceiling. GPT-6.1 Sol's index tops out at 52. Opus 5.5 scores 51 at its default, and only it has the 54, 56 and 58 above[2].
- Large requests cost no extra. The GPT models bill the whole request at higher rates above 272K input tokens; Opus 5.5's 1M context has one price. DataCamp's advice is to pick Opus 5.5 for long agentic coding or requests beyond 272K[1].
- Cheaper than Fable 5.2 on daily work. Its list price is half of Fable 5.2's, and on ComputingForGeeks's three infrastructure tasks the two were level on correctness with Opus 5.5 costing less than half. Not necessarily on hard problems: on FrontierCode 2.0, Fable 5.2 at High (45.3%, $6.72 per task) matches Opus 5.5 at Max (45.5%, $12.82). Hence its advice to start on Opus 5.5 for most work and keep Fable 5.2 for where its lead is real[3].
Good for
- Daily development, refactoring and code review in Claude Code
- Long agent runs that coordinate several tools
- Knowledge-work documents and long-context analysis
Less suited to
- High-volume, simple, cost-sensitive work (use GPT-6.1 Sol or something lighter)
- Short, well-specified scripts: a cheaper model gives the same result
- Chat clients that only speak the OpenAI format (the default group has no Chat Completions endpoint for it)
For the tier above, see the Claude Fable 5.2 review; for the cheaper alternative, the GPT-6.1 Sol and Terra review.
How to call it#
In Claude Code:
export ANTHROPIC_BASE_URL="https://moerouter.com"
export ANTHROPIC_AUTH_TOKEN="$MOEROUTER_API_KEY"
unset ANTHROPIC_API_KEY
claude --model claude-opus-5-5
With the Anthropic SDK:
import os
from anthropic import Anthropic
client = Anthropic(
base_url="https://moerouter.com",
api_key=os.environ["MOEROUTER_API_KEY"],
)
message = client.messages.create(
model="claude-opus-5-5",
max_tokens=4096,
messages=[{"role": "user", "content": "Refactor this module without changing its public interface, and give the reason for each step"}],
)
print(message.content[0].text)
The most common mistake is calling it from a chat client over Chat Completions: in the default group claude-opus-5-5 is offered on the Messages endpoint only, and the answer is a 404 that says to call /v1/messages. See endpoints and protocols. The docs are in Chinese for now.
Frequently asked questions#
Claude Opus 5.5 or Fable 5.2?#
Opus 5.5 for daily development. Its list price is half of Fable 5.2's and it scores higher on Bito's coding benchmark. Fable 5.2 is worth it only for agent runs that last hours and for problems other models cannot get through.
How much does the Claude Opus 5.5 API cost?#
List price per 1M tokens is $4 input and $20 output, with cache reads at $0.20 and no surcharge across the 1M context. MoeRouter's reference prices are in the table above; the three groups cost differently, and accounts are charged in Chinese yuan.
How does Opus 5.5 compare with GPT-6.1 Sol?#
Up to an index score of 51, Sol is cheaper at every level. Above 51 only Opus 5.5 goes, at $1.82 to $5.98 per task. The usual set-up is Sol by default and Opus 5.5 when a task fails or is plainly hard.
Why does a chat client get a 404 for claude-opus-5-5?#
Most chat clients use the Chat Completions endpoint, and in the default and stable groups claude-opus-5-5 is offered on Anthropic Messages only. Set the client's API type to Anthropic, or use a key of the ccmax-cheap group.
How do I use Opus 5.5 in Claude Code?#
Set ANTHROPIC_BASE_URL to https://moerouter.com and ANTHROPIC_AUTH_TOKEN to your MoeRouter key, and start with --model claude-opus-5-5. The Claude Code guide has the full steps.
How high should I set Opus 5.5's reasoning effort?#
The default, medium, already scores 51 on the index. Higher levels take a task from $1.82 to $5.98 for 54 to 58 points, which pays only when the task is really hard.
Sources
The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.
- GPT-6 Sol vs Claude Opus 5.5: Benchmarks, Price, and a Test
- Claude Opus 5.5 vs GPT-6.1 Sol: Level at Default Effort, No Contest Above It
- Claude Fable 5.2: Benchmarks, Pricing and Real Tests
- GPT-6.2 Sol and Astra benchmark: how they compare to Claude Opus 5.5, Fable 5.2 and GPT-6.1 Sol for coding
- GPT-6.1 Sol vs. Claude Opus 5.5: Which Model to Use
- Opus 5.5、GPT-6 Sol首发实测,Claude再次统治世界?! (video, in Chinese)
- Claude Opus 5.5、GPT6 Sol、Grok 4.7实测:谁更值? (video, in Chinese)
- 能否挑战天花板? Gemini 4 Argon 对比 Claude Opus 5.5 实测 (video, in Chinese)
- Claude Fable 5.2实测:94分刷爆了我们的私有题库,但我真用不起了! (video, in Chinese)
Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.