# Claude Opus 5.5 vs GPT-6.1 Sol and GPT-6 Sol: When Opus Pays Off

> Against GPT-6.1 Sol, Opus 5.5 costs exactly 2x per token under 272K input and stays pricier above. Opus scores higher; pay only where it passes more tasks.

- URL: https://blog.laozhang.ai/en/posts/claude-opus-5-5-vs-gpt-6-sol
- Published: 2026-09-23
- Updated: 2026-10-01
- Author: LaoZhang AI Team (https://blog.laozhang.ai/en/about)
- Topic: Model Comparisons
- Tags: Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Sol, Coding Agents, AI Model Comparison, API Pricing

---
GPT-6.1 Sol, released on September 29, 2026 as the upgrade to GPT-6 Sol, is now the Sol to weigh against Claude Opus 5.5. For the same tokens in a request under 272,000 input tokens, Opus 5.5 costs exactly twice as much in every billing column, cache reads included: GPT-6.1 Sol cut cached input to $0.10 per million, while Opus charges $0.20. Above 272K, OpenAI reprices the entire Sol request and the gap narrows, but for the same tokens Opus stays more expensive at any mix. Opus 5.5 still scores higher on Artificial Analysis' Intelligence Index at every effort level except low, where the two tie.

List rates are only half the story. On Artificial Analysis' task set at matched effort, Opus cost 4.2 to 8.9 times as much per task as GPT-6.1 Sol, and 6.4 times at the default medium effort, because it spent far more tokens. The number that should decide your default is the cost of getting one task accepted, including retries and review time. GPT-6 Sol (`gpt-6-sol`) has no announced shutdown date, so its figures stay below as a reference. All prices are USD per million tokens on each vendor's standard tier, as of September 30, 2026 for OpenAI and September 25, 2026 for Anthropic.

## The rates that actually differ

| Per 1M tokens | Claude Opus 5.5, any request size | GPT-6.1 Sol, up to 272K input | GPT-6.1 Sol, over 272K input | GPT-6 Sol (reference), up to / over 272K |
| --- | --- | --- | --- | --- |
| Fresh input | $4 | $2 | $4 | $2 / $4 |
| Cache write | $5 (5-minute) or $8 (1-hour) | $2.50 | $5 | $2.50 / $5 |
| Cache read | $0.20 | $0.10 | $0.20 | $0.20 / $0.40 |
| Output, including reasoning | $20 | $10 | $15 | $10 / $15 |

Sources: [Claude API pricing](https://platform.claude.com/docs/en/about-claude/pricing), [OpenAI API pricing](https://developers.openai.com/api/docs/pricing), and the model pages for [GPT-6.1 Sol](https://developers.openai.com/api/docs/models/gpt-6.1-sol) and [GPT-6 Sol](https://developers.openai.com/api/docs/models/gpt-6-sol). OpenAI's pricing table now lists only GPT-6.1 Sol; the GPT-6 Sol page still shows its rates and points to 6.1 as "the newer Sol model."

Two rules shape everything that follows. Anthropic bills Opus 5.5 at the same rate across its full 1M-token window; in its own words, "a 900k-token request is billed at the same per-token rate as a 9k-token request." OpenAI does the opposite for both Sol models: prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output "for the full request," not just for the tokens past the line.

Both vendors offer roughly a million tokens of context (1,050,000 for Sol, 1M for Opus) and 128,000 output tokens. Both run with reasoning. GPT-6.1 Sol's `reasoning.effort` goes from `low` to `max` with `medium` as the default; unlike GPT-6 Sol, it does not accept `none` or `minimal`. Opus 5.5 keeps adaptive thinking on at all times, also defaulting to medium effort. If you are moving existing `gpt-6-sol` calls, [GPT-6.1 Sol: Pricing, Access, and What Breaks When You Switch](https://blog.laozhang.ai/en/posts/gpt-6-1-sol) covers the effort and tool-calling changes.

The usual modifiers move both bills in the same direction. Batch halves input and output on both, and OpenAI's Flex tier also halves Sol. Fast mode doubles both: GPT-6.1 Sol to $4 and $20 in short context, Opus to $8 and $40. Regional processing adds 10% on OpenAI for eligible models, and Anthropic's US-only inference adds 1.1x to every token category. OpenAI's built-in tools can carry per-call fees on top, such as $10 per 1,000 web search calls.

## Where the 2x gap holds and where it shrinks

Put your own billed token counts, in millions, into these formulas:

```text
U = fresh input   W = cache writes   R = cache reads   O = output (incl. reasoning)

Claude Opus 5.5                    = 4U + 5W + 0.2R + 20O
GPT-6.1 Sol, up to 272K            = 2U + 2.5W + 0.1R + 10O
GPT-6.1 Sol, over 272K             = 4U + 5W + 0.2R + 15O

GPT-6 Sol (reference), up to 272K  = 2U + 2.5W + 0.2R + 10O
GPT-6 Sol (reference), over 272K   = 4U + 5W + 0.4R + 15O
```

The Opus line uses 5-minute cache writes; 1-hour writes at $8 widen the gap further. Two shortcuts fall out of the GPT-6.1 Sol lines:

- **Under 272K, Opus costs exactly 2 × GPT-6.1 Sol for the same tokens.** Every term doubles, so a cache-heavy workload keeps the full 2x gap. Against GPT-6 Sol, cache reads cost $0.20 on both sides, so the ratio used to be 2 minus the share of Sol's bill spent on cache reads. That shortcut applies only to the older model.
- **Over 272K, Opus minus GPT-6.1 Sol equals 5O.** Input and cache rates are identical there, and only output differs ($20 against $15). The gap narrows as output shrinks relative to input, but it never goes negative. Against GPT-6 Sol the difference was 5O − 0.2R, so Opus became the cheaper model once cache reads exceeded 25 times output. With 6.1 Sol that crossover no longer exists.

These worked mixes use identical billed token counts on both sides, which real runs never produce. Treat them as a view of the rate cards, not as bills:

| Hypothetical workload | Tokens (millions) | Opus 5.5 | GPT-6.1 Sol | Opus ÷ 6.1 Sol | GPT-6 Sol (reference) | Opus ÷ 6 Sol |
| --- | --- | --- | --- | --- | --- | --- |
| A. Short prompts, no cache | U 0.1, O 0.02 | $0.80 | $0.40 | 2.00 | $0.40 | 2.00 |
| B. Agent loop of many short calls | U 0.1, R 2.0, O 0.05 | $1.80 | $0.90 | 2.00 | $1.10 | 1.64 |
| C. Cache-dominant loop | U 0.05, R 5.0, O 0.03 | $1.80 | $0.90 | 2.00 | $1.40 | 1.29 |
| D. One 800K-token request, mostly cached | U 0.05, R 0.75, O 0.02 | $0.75 | $0.65 | 1.15 | $0.80 | 0.94 |
| E. One uncached 300K-token request | U 0.3, O 0.02 | $1.60 | $1.50 | 1.07 | $1.50 | 1.07 |

Anthropic says cache reads "make up the majority of agentic and coding work costs." That is a statement about its own customers, but it matches how agents behave: each turn resends the growing conversation, most of it cached. Against GPT-6 Sol, that pattern pulled the two bills closer, as rows B and C show. Against GPT-6.1 Sol it doesn't, because Sol's cache reads are also half price. Caching still cuts both bills sharply; it just no longer changes the ratio. See [why cache misses get expensive](https://blog.laozhang.ai/en/posts/claude-code-cache-miss-token-costs) for how a falling hit rate plays out in Claude Code.

The 272K line applies per request, not per session. A long agent run made of short calls stays in Sol's cheaper band even if it bills millions of tokens in total. But once the conversation that each call resends passes 272K, every later call is priced at the long-context rates. That band is the only place where the list-rate gap drops below 2x, to 1.07 and 1.15 in rows D and E. OpenAI's wording ("2x input and cache rates") suggests that cached tokens count toward the threshold, so the safe reading is total prompt size, cached or not. OpenAI does not state it in so many words.

![Rate formulas for Claude Opus 5.5 and GPT-6.1 Sol: up to 272K input Opus costs exactly 2x for the same tokens, over 272K the difference is 5 times output, then retries, tool fees, and reviewer time divided by passing tasks give the cost per approved task](https://blog.laozhang.ai/posts/en/claude-opus-5-5-vs-gpt-6-sol/img/cost-per-approved-task.webp)

## Why measured cost per task runs 4x to 9x

Rate cards assume both models use the same tokens. They don't. [Artificial Analysis](https://artificialanalysis.ai/models/releases/comparisons/gpt-6-1-sol-vs-claude-opus-5-5) runs both models across the ten evaluations in its Intelligence Index v4.3.2 and prices each run from actual input, cache, reasoning, and answer tokens. As of September 30, 2026, at matched effort:

| Effort | Opus 5.5 index score | Opus 5.5 cost per index task | GPT-6.1 Sol index score | GPT-6.1 Sol cost per index task | Cost ratio |
| --- | --- | --- | --- | --- | --- |
| low | 42 | $0.55 | 42 | $0.13 | 4.2x |
| medium (default) | 51 | $1.34 | 48 | $0.21 | 6.4x |
| high | 54 | $1.82 | 50 | $0.32 | 5.7x |
| xhigh | 56 | $3.46 | 51 | $0.39 | 8.9x |
| max | 58 | $5.98 | 52 | $0.72 | 8.3x |

GPT-6 Sol, for comparison, peaked at 48 on the same index at max effort, for $1.05 per task. GPT-6.1 Sol matches that score at medium for a fifth of the cost.

Since no Opus rate is more than twice GPT-6.1 Sol's, a 6.4x cost gap at medium effort means Opus used at least about 3.2 times Sol's token volume on those tasks, weighted by price category. The index's own counts agree: at medium, Opus produced about 26,000 output tokens per task against 8,000 for 6.1 Sol, and at max 119,000 against 38,000. Both vendors bill reasoning as output, the most expensive column.

DataCamp's [single hands-on test](https://www.datacamp.com/blog/gpt-6-sol-vs-claude-opus-5-5) ran GPT-6 Sol before 6.1 shipped and shows the same pattern. In a one-attempt, single-file Tetris build at high effort in OpenCode, Opus needed fewer turns and fewer total tokens (107,958 against 120,226) yet cost $0.87 against Sol's $0.26. It spent 22,551 reasoning tokens to Sol's 8,123, all billed as output at twice Sol's rate. Sol's game scored 5.0 on DataCamp's rubric, Opus's 4.3. One task is an illustration, not a verdict.

Two more points keep this in proportion:

- **Compare quality levels, not effort labels.** GPT-6.1 Sol at max (52, $0.72) outscores Opus at medium (51, $1.34) for about 54% of the cost. At low effort the two tie on score, and Opus costs 4.2 times as much. In the other direction, Opus at xhigh and max (56 and 58) reaches scores that GPT-6.1 Sol does not reach at any effort on this index. If your work needs that level, Sol's lower price doesn't buy it.
- **Token counts don't transfer between vendors.** The two models use different tokenizers. Anthropic says its newer tokenizer, used by Claude 4.7 and later, produces about 30% more tokens for the same text than its previous one. No vendor publishes a Claude-to-OpenAI ratio, so the only honest comparison uses each API's own billed usage on the same task.

## Reading the scores without the usual traps

**Check which Sol is in the chart.** Anthropic's launch table for [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) compares it with GPT-5.6 Sol, two versions back. Its Terminal-Bench 4.0 line (66.4% for Opus at xhigh against 37.3%) and CursorBench 4.0 line (57.8% against 41.7%) say nothing about either GPT-6 model. The GPT figures there are "as reported by OpenAI," and Opus results mostly use max effort. Sol charts and tests published between September 22 and 29, 2026 show GPT-6 Sol, which scores lower than 6.1 Sol at every effort on Artificial Analysis' index. [GPT-6.1 Sol vs GPT-6 Astra vs GPT-6 Sol: When Astra Is Worth It](https://blog.laozhang.ai/en/posts/gpt-6-1-sol-vs-gpt-6-astra-vs-gpt-6-sol) sets the two Sols side by side.

**Match effort levels.** Every row of the table above moves by several points between efforts. A chart that pairs Sol at xhigh with Opus at medium is comparing settings, not models.

**Know who picked the benchmark.** OpenAI's [GPT-6.1 Sol announcement](https://openai.com/index/introducing-gpt-6-1-sol/) reports that on GDP.pdf, 6.1 Sol scored higher than Opus 5.5 with fallback at every tested reasoning setting, for less than half the cost per task. On AutomationBench 1.0.6 at medium effort, it led Opus by 2.2 points at about a third of the cost. On Terminal-Bench Science 0.1 at max effort, its average cost per task was $5.47 against $23.21 for Opus, though GPT-6 Astra still ranked first there. OpenAI chose these evaluations and took the Opus results from public reports, while Artificial Analysis' ten-evaluation index still puts Opus ahead overall. Both can hold: 6.1 Sol wins some task types for a fraction of the cost, and Opus leads on the aggregate. Anthropic's own "costs 40% less to run" claim is measured against Opus 5, not Sol. The company also writes that benchmark margins "have become a less reliable guide to real-world differences," which is a good reason to run your own tasks.

**Speed splits two ways.** As of September 30, 2026, Artificial Analysis measures GPT-6.1 Sol at 59 to 66 output tokens per second across efforts, against 72 for Opus at low and medium and 92 at max. Sol answers sooner, though: its time to first answer token was 2.12 seconds at low effort and 5.72 seconds at medium, against 12.72 and 22.95 seconds for Opus. Because Opus writes more tokens, a whole index task took about 136 seconds on average for 6.1 Sol at medium and about 219 seconds for Opus. These values are re-measured over time, so older snapshots differ. Anthropic lists Opus 5.5's latency as "Moderate."

## In Claude Code and Codex, plan limits matter more than token prices

On a subscription, the rate card above does not apply directly. Claude plans and the Claude API are separate products with separate bills, and so are ChatGPT plans and the OpenAI API. OpenAI says as much: API token prices "are separate from subscription usage; don't use them to estimate included tasks."

Where each model is available:

- **GPT-6.1 Sol** is in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise, and Edu plans. Enterprise and Edu keep it off until an administrator enables it, and Free and Go are not included. It is not in regular ChatGPT Chat. In the CLI, `codex --model gpt-6.1-sol` selects it. GPT-6 Sol stays available in Work and Codex, and GPT-5.6 models remain during the rollout. [OpenAI's model page](https://learn.chatgpt.com/docs/models) notes that selecting a model does not grant access to it.
- **Claude Opus 5.5** is in the Claude apps on Pro, Max, Team, and Enterprise, in Claude Code from version 2.1.280, on the Claude API, and on AWS, Google Cloud, and Microsoft Azure. The full ID `claude-opus-5-5` is safer than the `opus` alias, whose target depends on the provider.

A widely shared complaint asks why Sol, cheaper per token, seems to drain a $20 Codex plan faster than Opus drains Claude Code. Public numbers can only answer part of it. [Codex's pricing page](https://learn.chatgpt.com/docs/pricing) estimates 15–160 local GPT-6.1 Sol messages per five hours on Plus, against 15–150 for GPT-6 Sol. It stresses that these are "not fixed message limits" and that the model and task change the count, and it adds that weekly limits may also apply. Pro now comes at $100, $200, or $500 a month, and Pro plans "currently have no five-hour limit." Anthropic raised five-hour limits on its paid plans with Opus 5.5 but published no message estimate. Each vendor sets its plan allowances per product, not from its API rates, so token prices cannot predict which plan lasts longer. Your usage dashboard can.

For the plan-level comparison, see [Claude Code vs Codex Cost: Plans, Limits, and the Real API Bill](https://blog.laozhang.ai/en/posts/claude-code-vs-codex). What happens when Plus runs out is covered in [ChatGPT Plus Codex Weekly Limits: Do Credits Add More Usage?](https://blog.laozhang.ai/en/posts/chatgpt-plus-codex-weekly-limit-credits), and the Opus 5.5 plan changes in [Claude Opus 5.5 Pricing and Limit Reset](https://blog.laozhang.ai/en/posts/claude-opus-5-5-pricing-limit-reset).

## A trial that ends in a decision

Cost per accepted task is the number that settles it:

```text
cost per accepted task =
  (API charges for every attempt, including retries and tool fees
   + reviewer hours × hourly rate)
  ÷ number of tasks that pass your acceptance criteria
```

![Six-step trial for Claude Opus 5.5 and GPT-6.1 Sol: pick real tasks, hold everything else constant, record run conditions, read tokens from each API usage field, grade blind and time the review, then price every attempt to get the cost per accepted task](https://blog.laozhang.ai/posts/en/claude-opus-5-5-vs-gpt-6-sol/img/acceptance-trial.webp)

1. **Pick real tasks and write the pass criteria first.** Ten to twenty tasks from your own backlog work better than a public benchmark: a bug fix with a failing test, a refactor, a research summary. Decide before any run what "accepted" means, such as tests pass, no behavior change, or a reviewer signs off without edits.
2. **Hold everything else constant.** Use the same prompt, repository snapshot, tools, and harness, and set a retry budget, for example two retries per task. Set effort explicitly on both sides; the default is medium for each, and GPT-6.1 Sol's lowest setting is `low`. Comparing mismatched levels is the most common error.
3. **Record the run conditions.** Log the model ID (`gpt-6.1-sol` and `gpt-6-sol` are different models), effort, cache policy, service tier (Standard, Batch, Flex, or Fast), regional settings, and date for every run.
4. **Take token counts from each API's usage fields, not from estimates.**

| Billing column | Claude Messages API `usage` | OpenAI Responses API `usage` |
| --- | --- | --- |
| Fresh input (U) | `input_tokens` | `input_tokens` minus cached and cache-write tokens |
| Cache writes (W) | `cache_creation_input_tokens` | `input_tokens_details.cache_write_tokens` |
| Cache reads (R) | `cache_read_input_tokens` | `input_tokens_details.cached_tokens` |
| Output (O) | `output_tokens` | `output_tokens` |

OpenAI's `input_tokens` is the request's total input. Any Sol request above 272,000 must be priced entirely at the long-context rates. On the Claude side, total input is the sum of the three input fields.

5. **Grade blind if you can, and time the review.** Record pass or fail, retries used, and reviewer minutes per task. A rejected result costs a rerun plus the time someone spent finding the problem.
6. **Price every attempt with the formulas above**, add tool fees, and divide by accepted tasks.

Reviewer time often outweighs the API bill. In a hypothetical case, Sol costs $0.40 per attempt and Opus $1.50. Each rejected result also costs 15 minutes of an $80-per-hour reviewer, which is $20. Across ten tasks Opus adds $11 in API charges, so one fewer rejection pays for it with room to spare. If both models pass the same tasks, Opus's premium buys nothing.

## When to change your default

- **Stay on GPT-6.1 Sol** when output makes up much of the bill, results are cheap to check (tests, schemas, linters), or time to first token matters. At list rates it is cheaper at every request size, and at similar acceptance rates it is the better default. If you still run `gpt-6-sol`, move to 6.1 first: same input and output prices, half-price cache reads, and higher index scores at every effort.
- **Switch to Claude Opus 5.5** when your trial shows it passing tasks that GPT-6.1 Sol fails even at xhigh or max, or when reviewer time per task dwarfs the API charges. Long, cache-heavy requests above 272K are no longer a price reason on their own: Opus still costs more there, but only by 5 × the output cost, 1.15x in row D. Near parity, even a small gain in acceptance covers it.
- **Raise Sol's effort before switching.** On Artificial Analysis' numbers, GPT-6.1 Sol at max outscores Opus at medium for about half the cost per task. If a higher effort closes the acceptance gap, that is the cheaper fix.
- **Split by task type** if the trial is mixed. Route checkable, high-volume work to Sol and send the long, hard-to-verify jobs, especially those that need Opus's xhigh or max quality, to Opus.

If you are also weighing other models in either lineup, [GPT-6 Sol vs Terra vs Luna vs Astra](https://blog.laozhang.ai/en/posts/gpt-6-sol-vs-terra-vs-luna-vs-astra) covers OpenAI's options, including Luna for cheap volume. [Claude Fable 5.1 vs Opus 5.5](https://blog.laozhang.ai/en/posts/claude-fable-5-1-vs-opus-5-5) covers when to step up from Opus.

## Sources

External pages this guide links to, in the order they appear. Last updated 2026-10-01.

- [Claude API pricing](https://platform.claude.com/docs/en/about-claude/pricing) (platform.claude.com)
- [OpenAI API pricing](https://developers.openai.com/api/docs/pricing) (developers.openai.com)
- [GPT-6.1 Sol](https://developers.openai.com/api/docs/models/gpt-6.1-sol) (developers.openai.com)
- [GPT-6 Sol](https://developers.openai.com/api/docs/models/gpt-6-sol) (developers.openai.com)
- [Artificial Analysis](https://artificialanalysis.ai/models/releases/comparisons/gpt-6-1-sol-vs-claude-opus-5-5) (artificialanalysis.ai)
- [single hands-on test](https://www.datacamp.com/blog/gpt-6-sol-vs-claude-opus-5-5) (datacamp.com)
- [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) (anthropic.com)
- [GPT-6.1 Sol announcement](https://openai.com/index/introducing-gpt-6-1-sol/) (openai.com)
- [OpenAI's model page](https://learn.chatgpt.com/docs/models) (learn.chatgpt.com)
- [Codex's pricing page](https://learn.chatgpt.com/docs/pricing) (learn.chatgpt.com)