- Home
- Learn
- Deep Dives
- Claude Opus Models Compared: Opus 4.5 to Opus 5.5
Claude Opus Models Compared: Opus 4.5 to Opus 5.5
Every Claude Opus version side by side: release dates, context, API prices and what a coding session costs. Opus 4.5 to 5.5, checked 16-11-2026.
Published 16-11-2026. Last verified on 16-11-2026 against Anthropic's model overview, pricing and deprecation pages. We update this page when a new Opus version ships and re-check it monthly.
Changelog: 16-11-2026, first version, covering Opus 4.5, 4.6, 4.7, 4.8, 5 and 5.5.
Quick answer: which Claude Opus should you use today?
This comparison is for developers and engineering leads who run Claude Code or their own agents on Claude Opus and have to decide which version to pin, and what that choice costs per working session. On 16-11-2026 Anthropic lists six Opus versions as active in the API: Opus 4.5, 4.6, 4.7, 4.8, 5 and 5.5. Claude Opus 5.5, released on 22-09-2026, is the one we would put almost every team on: it is the newest, it has the 1M-token context window and at $4 input and $20 output per million tokens it is the cheapest Opus on Anthropic's current price list. Opus 4.6 to Opus 5 all cost $5 and $25. Opus 4.5 is still callable but has the smaller 200K context window and may be retired from 24-11-2026. The older Opus 4 and 4.1 are already retired.
How we compared the Opus models
Our method in three lines: every date, price and context figure comes from Anthropic's own pages as we read them on 16-11-2026; every cost on this page is our calculation on those prices, with the assumptions written out; and benchmark numbers are labelled as vendor-reported, because we did not run them.
Our gate for this page: a version only enters the main table if Anthropic's deprecation page lists it as active in the API on the date above. Previews and retired models sit in a separate block. The gate removed three things. Opus 4 and Opus 4.1 left the main table because they were retired on 15-06-2026 and 05-08-2026. Fast mode on Opus 5 stayed out of the price columns because Anthropic calls it a research preview. And the long-context surcharge from the Opus 4.6 launch post, for prompts over 200K tokens, is not on this page because the current pricing page no longer shows it for Opus.
We also had no model access in this production run, so the cost example below is calculated, not measured. Where we would like a measured number, we say so.
What is Claude Opus for?
Opus is Anthropic's tier for the hardest work. The models overview describes Opus 5.5 as the model "for long-running agentic coding and knowledge work", Sonnet 5.5 as the best combination of speed and intelligence and Haiku 5.5 as the model for high-volume, latency-sensitive jobs such as classification and routing. Above Opus sits a separate line, Fable, which the same page recommends only when evals on Opus 5.5 at higher effort still fall short.
We think Opus earns its price on three kinds of jobs: multi-file changes where one wrong assumption breaks the build, agents that run for an hour or more without a human checking each step and analysis where the model has to hold a large document set in context. For those, a stronger first attempt saves more in review time than the extra tokens cost. If your team builds multi-step agents on Opus, the agents and orchestration collection has skills for planning, handoffs and run loops.
All three top-ranking pages for "claude opus" agree on the same core: Opus is the flagship, it is aimed at coding and reasoning and it handles long context. Where they differ is the version they describe, which is why we put every version in one table.
Every Opus version side by side
The main table holds the versions Anthropic lists as active on 16-11-2026. Release dates come from Anthropic's launch posts; context, output and prices from the models overview and the pricing page. All prices are per million tokens, API list price, vendor-reported.
| Version | API ID | Released | Context | Max output | Input | Output | Earliest retirement |
|---|---|---|---|---|---|---|---|
| Opus 5.5 | claude-opus-5-5 | 22-09-2026 | 1M | 128K | $4 | $20 | 22-09-2027 |
| Opus 5 | claude-opus-5 | 24-07-2026 | 1M | 128K | $5 | $25 | 24-07-2027 |
| Opus 4.8 | claude-opus-4-8 | 28-05-2026 | 1M | 128K | $5 | $25 | 28-05-2027 |
| Opus 4.7 | claude-opus-4-7 | 16-04-2026 | 1M | 128K | $5 | $25 | 16-04-2027 |
| Opus 4.6 | claude-opus-4-6 | 05-02-2026 | 1M | 128K | $5 | $25 | 05-02-2027 |
| Opus 4.5 | claude-opus-4-5-20251101 | 24-11-2025 | 200K | 64K | $5 | $25 | 24-11-2026 |
The last column is the "not sooner than" date from the model deprecations page. We would not start a new project on Opus 4.5: its earliest retirement is eight days after this page was checked.
Retired, no longer callable: Opus 4.1 (claude-opus-4-1-20250805, $15 and $75, retired 05-08-2026) and Opus 4 (claude-opus-4-20250514, $15 and $75, retired 15-06-2026). Anthropic's recommended replacement for the older Claude 3 Opus is Opus 4.8. If you still see either 4.x ID in a script, that call now fails.
What changed from Opus 4.5 to Opus 5.5?
Each step changed one thing that matters for a team's bill or workflow. In release order, with what Anthropic said at launch:
- Opus 4.5 (24-11-2025). It sits at the start because it cut the Opus price from $15 input and $75 output to $5 and $25, which made Opus a daily model instead of a special-occasion one. Anthropic's launch post says that at medium effort it matched Sonnet 4.5's best SWE-bench Verified score with 76% fewer output tokens.
- Opus 4.6 (05-02-2026). Second because it brought the 1M-token context window and 128K output to Opus. Anthropic reported 76% on the 8-needle 1M variant of MRCR v2, against 18.5% for Sonnet 4.5.
- Opus 4.7 (16-04-2026). Third because of a change you will see on the invoice rather than in quality: per the pricing page, Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text. Same list price, more tokens.
- Opus 4.8 (28-05-2026). Fourth: Anthropic's post shows its gains mainly in a chart image, and we could not read absolute numbers from the text. Fast mode launched on it and is now listed as deprecated for this version.
- Opus 5 (24-07-2026). Fifth: it became the default model on Claude Max. Anthropic claims, without numbers in the text, that it more than doubles Opus 4.8 on Frontier-Bench v0.1.
- Opus 5.5 (22-09-2026). Last and newest: lower price, thinking always on (adaptive, default effort medium in the API) and, per Anthropic, Fable 5.1-level performance on most work at 40% lower running cost than Opus 5.
The upgrade from 4.6 to 4.7 is the one we would test most carefully. Your prompts do not change, but token counts do, and an agent that sat just under a context or budget limit can now cross it. Debugging those regressions is a job for the testing and debugging collection.
What does Claude Opus cost?
On the API, every active Opus except 5.5 costs $5 per million input tokens and $25 per million output tokens. Opus 5.5 costs $4 and $20. Those are list prices from the pricing page on 16-11-2026. Three discounts change the real number far more than the version choice:
| Option | Opus 4.5 to Opus 5 | Opus 5.5 |
|---|---|---|
| Cache hit (input read from cache) | $0.50 | $0.20 |
| Cache write, 5 minutes | $6.25 | $5 |
| Batch input and output | $2.50 and $12.50 | $2 and $10 |
| Fast mode input and output | $10 and $50 (Opus 4.8 and 5) | $8 and $40 |
Opus 5.5 reads cached input at 0.05 times the base price, against 0.1 times for other models, which is why we think caching matters even more on 5.5 than on the older versions. Requests that require US-only inference on Claude 4.6 and later carry a 1.1 times multiplier on all token prices.
On subscription plans you do not pay per token. Anthropic's Opus 5.5 page says the model is available to Pro, Max, Ultra, Team and Enterprise users. Anthropic does not publish an Opus message count per plan on those pages, so we will not give one; our guide to Claude Code usage limits explains how the weekly caps work. For the full API price list across models, see Claude API pricing explained.
What does a 50-message coding session cost on each Opus?
Here is our worked example. Assumptions: a 50-message Claude Code session where each message sends about 30,000 input tokens (code, history, tool results) and returns about 2,000 output tokens, so 1.5 million input and 100,000 output tokens counted with the old tokenizer. For Opus 4.7 and later we add 30% to both, following Anthropic's tokenizer note. The cached column assumes 90% of input is read from cache and ignores cache-write costs. These are calculations on vendor list prices, not measurements.
| Version | No caching | 90% cache hits |
|---|---|---|
| Opus 4.5 and 4.6 | $10.00 | $3.93 |
| Opus 4.7, 4.8 and 5 | $13.00 | $5.10 |
| Opus 5.5 | $10.40 | $3.73 |
Two things stand out to us. First, moving from Opus 4.6 to Opus 5 at the same list price makes this session about 30% more expensive purely through token counting. Second, caching cuts every row by more than half, which is a bigger lever than any version switch. If you want to know what your own sessions cost instead of trusting our assumptions, Agent Cost Scan, free, reads your agent logs and reports cost per run, the most expensive model and ranked fixes. Cut Agent Token Costs: Caching, Routing and Budget Guards, $5, then applies the caching and routing. Our guide to reducing Claude Code token usage covers the habits, and the LLM ops collection has more skills for cost control.
What do the benchmarks show?
Less than most pages suggest, and we would rather say that than invent a table. Anthropic publishes SWE-bench Verified results for Opus 4.5 to Opus 5 mainly as chart images or relative claims in its launch posts, and the swebench.com leaderboard did not return its rows when we fetched it on 16-11-2026. So we do not give a SWE-bench Verified score per version here.
What Anthropic does publish in text is the comparison table in the Opus 5.5 launch post, all at max effort and all vendor-reported:
| Benchmark | Opus 5.5 | Opus 5 | Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 52.3% | 55.8% |
| FrontierCode v1.1 (Main) | 54.4% | 48.0% | 50.3% |
| CursorBench 4.0 | 57.8% | 46.6% | 51.8% |
| OSWorld 2.1 (partial) | 81.8% | 74.0% | 80.7% |
For Opus 4.5, the launch post gives a relative claim: at its highest effort it beat Sonnet 4.5 on SWE-bench Verified by 4.3 percentage points while using 48% fewer tokens.
Vendor tables are a starting point, not a verdict. If you want independent data inside the same session where you pick the model, 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 coding leaderboard from LMArena, from $0.013 per call (starting prices read on 16-11-2026). You do not need an account or API key at either provider, the price is shown before the call runs and a failed call is free. To test models on your own tasks instead, Agent Evals with Real Statistics, $12, sets up golden datasets, confidence intervals and paired comparisons. More in the research and analysis collection.
When is Opus worth it over Sonnet?
Our routing rule: Opus for work where a wrong first answer costs a person more than ten minutes to find and fix, Sonnet for everything else. That puts architecture changes, migrations across many files, security review and long unattended agent runs on Opus, and everyday edits, tests and documentation on Sonnet. We hold this as a rule of thumb, not a measured threshold; we have not run the same task set on both for this page.
The cost gap is smaller than it used to be: Opus 5.5 is 20% cheaper per token than Opus 4.5 to 5. For a team doing mostly hard work, we think routing everything to Opus 5.5 can be simpler than maintaining two paths. For mixed workloads, Route Each Task to the Right Model, $12, sends easy tasks to a cheaper model and escalates only when a check fails. Our best Claude model for coding guide goes deeper on the coding side.
Code review is the clearest Opus job in our view, because it is the place where one missed issue gets expensive later. AI Code Reviewer and Security Analyzer, free, runs an automated review and security pass on changes. The code quality and review collection has the rest.
Opus 5 vs Fable 5: which one?
For most teams, neither: we would take Opus 5.5. Between the two in the question, Fable 5 costs $10 input and $50 output per million tokens, double Opus 5, and Anthropic positions Fable for demanding reasoning and long-horizon agentic work. Anthropic's own table puts Opus 5.5 ahead of Fable 5.1 on all four benchmarks above, at less than half the list price of Fable 5.1. The newest Fable, Fable 5.2, costs $8 and $40. We would only move a workload to Fable after evals on Opus 5.5 at higher effort fall short, which is also what Anthropic's overview advises. Our Claude Fable 5 skills guide covers that line in detail.
Which skills benefit most from Opus?
Skills that make the model reason over a lot of material in one pass gain the most from Opus, because that is where a stronger model changes the outcome rather than the speed. Three types stand out to us:
- Multi-part audits. Full Audit Pack: 8 code audits, one score, memory, $19, runs eight audits on one core and produces a single scored report. First on our list because cross-checking eight audit results is exactly the kind of long reasoning Opus is priced for.
- Context-heavy agents. Agent Context Engineering: Prompts, Caching and Cost, $7, structures prompts, cache order and compaction. Second because it pays off most on the 1M-token models, where a badly ordered context is expensive.
- Documentation from large codebases. Third: writing accurate docs means reading a lot of code first. The technical documentation collection has skills for it.
Skills that run short, repetitive steps, such as formatting or tagging, rarely need Opus. We would run those on Sonnet or Haiku.
We sell skills and pay-per-call tools that your agent uses through one MCP connection, and we name ours in the sections above. No vendor paid to be listed.
Your next step
Pin your projects to claude-opus-5-5 unless you have a measured reason not to, move anything still on Opus 4.5 before 24-11-2026 and turn on prompt caching. To pull benchmark and cost data into the same session where you choose, connect your agent once through Agensi MCP. We will update this table when the next Opus ships.
Written by Samuel Rose, Founder of Agensi. Next scheduled review: 16-12-2026.