Products & Services · · Team LATENT
GPT-6.1 Sol launches into a cost contest with Opus 5.5
At half Opus 5.5's standard API rates, GPT-6.1 Sol is a contender for coding and computer use. But the price gap narrows on long inputs, and benchmarks use different conditions. We examine both companies' documentation to assess the trade-offs.
Ask an AI to fix code, read documents, and operate a screen when needed. For repeated work like this, accuracy on a single question is only part of the picture. Can it finish the job? How long will you wait? What does it cost, including retries?
OpenAI's GPT-6.1 Sol, released September 29, is an update aimed squarely at this kind of work. OpenAI describes it as offering capabilities close to the higher-tier GPT-6 Astra at lower cost. Claude Opus 5.5, announced September 22, likewise promises greater efficiency than its predecessor on extended development tasks and knowledge work. OpenAI changelog; GPT-6.1 Sol announcement; Opus 5.5 announcement
For those choosing an AI to carry out work with cost in mind, GPT-6.1 Sol is a strong candidate to compare with Opus 5.5. Its standard API input, output and cache-read unit prices are each half those of Opus 5.5. The available evidence does not, however, establish an across-the-board reversal in performance. There are also exceptions to this price gap for long inputs.
Key points
- GPT-6.1 Sol succeeds GPT-6 Sol. It is not simply Astra sold more cheaply, but a separate model that changes the balance of capability and cost for complex work.
- Both companies compete in code modification, computer use, and business workflows. Benchmark versions, reasoning effort, execution software, and model fallback conditions differ, so published results cannot be combined into one definitive ranking.
- Sol's rates rise above 272,000 input tokens. Selection should account for total consumption, completion time, and human corrections as well as per-token prices.
Sources checked September 30, 2026. Announcement dates follow the official documents. This is an analysis of public information, not a LATENT review measuring both models under identical conditions.
Contents
- Why GPT-6.1 Sol belongs in a comparison with Opus 5.5
- Coding comparisons need matching tests and execution conditions
- Computer-use and business-workflow tests differ in tasks and scoring
- The API price gap narrows on long inputs
- Distinguish response speed from time to finish the work
- Availability and considerations when delegating actions to AI
- Choose for your own workload
Why GPT-6.1 Sol belongs in a comparison with Opus 5.5
Despite their similar names, GPT-6 Sol, GPT-6.1 Sol, and GPT-6 Astra should be distinguished. GPT-6 Sol and Luna became available September 22; GPT-6.1 Sol followed the next week. Astra remains a separate, higher-tier model. OpenAI's model documentation recommends comparing Sol with Astra on your own tasks to assess the quality and cost difference. Official changelog; GPT-6.1 Sol specifications
That positioning helps explain the competition with Opus 5.5. Anthropic presents Opus 5.5 as suited to extended agentic coding and knowledge work, and as a starting point for everyday workloads. The models overlap in tasks that involve using tools and completing multiple steps, rather than simply answering questions. Claude model overview
OpenAI's announcement includes cost-performance comparisons with Opus 5.5 on GDP.pdf, for understanding PDFs, and AutomationBench, for business workflows. The choice of comparisons shows that Opus is a competitive reference point. It does not establish development motives such as rushing the release to beat Opus.
If price alone is the criterion, Opus is not the only alternative. The current Claude Sonnet 5.5 also costs $2 per million standard input tokens and $10 per million output tokens, matching Sol. The useful question in comparing with Opus is where paying more still makes sense for complex work, not simply whether a cheaper AI exists. Anthropic pricing
Coding comparisons need matching tests and execution conditions
Software development demands more than generating a few lines of code. A model must understand an existing structure, edit multiple files, and continue until tests pass. DeepSWE evaluates extended development tasks of this kind. Its provider, Datacurve, says it creates tasks without copying existing patches and tests software behavior rather than the shape of the code. DeepSWE
According to OpenAI, GPT-6.1 Sol matched Astra on DeepSWE v1.1 at roughly one-fifth the cost. It also exceeded GPT-6 Sol's highest score by 6.4 points with less reasoning effort and lower cost. The company's comparison suggests that more work which required higher budgets or reasoning effort with the previous Sol may now be handled at a lower cost. GPT-6.1 Sol announcement
Opus 5.5's system card reports 74.2% on DeepSWE v1.1, averaged across five runs. However, the Datacurve public leaderboard retrieved for this article did not contain both models' new results, so we could not establish a direct comparison under matching execution conditions. System card, page 175; DeepSWE public leaderboard
Anthropic's announcement reports 66.4% on Terminal-Bench 4.0 and 57.8% on CursorBench 4.0, which includes tasks spanning multiple files. But it does not provide GPT-6.1 Sol results. Figures for GPT-5.6 Sol or Astra in a comparison published before Sol 6.1's release cannot be substituted for the new model's performance. Opus 5.5 performance table
The conditions also differ. Anthropic's table generally uses max reasoning effort, but its Terminal-Bench 4.0 result uses xhigh for Opus and OpenAI's published high result for Astra. Opus's 66.4% also has a standard error of ±2.6 percentage points. According to the system card, Opus ran 66 tasks five times each in Claude Code's --bare mode, for a total of 330 attempts. Safety-driven switching to another model affected the results in 10% of attempts. System card, page 178
The software that supplies tools and execution procedures also influences the result. In the FrontierCode evaluation published by Anthropic, for example, Claude ran in Claude Code while GPT ran in Codex CLI. A model name and a score alone conceal that difference. System card, pages 175–177
OpenAI also states that it evaluated its own models in research environments or through the API and used public reports for competitors. This does not mean the prompts and available tools were the same as in everyday Codex or Claude Code use. These materials show improvement over the previous Sol, but its relative performance against Opus needs separate assessment. They do not establish that Sol wins at coding or that Opus is better.
Computer-use and business-workflow tests differ in tasks and scoring
Computer use has improved as well. OpenAI says GPT-6.1 Sol scored seven points higher than GPT-6 Sol at maximum reasoning effort on OSWorld 2.0's offline tasks, at less than half the cost, and narrowed the gap to Astra to 2.1 points. These are scores including partial credit on the August 8, 2026 task set. OpenAI's computer-use evaluation
Anthropic's announcement reports 81.8%, including partial credit, for Opus 5.5 on OSWorld 2.1. Its system card labels the same 81.8% result 2.0 and says it used the September 10 task set. The naming discrepancy remains unresolved between the documents, but at a minimum, that task set differs from OpenAI's August 8 version. Announcement; system card, pages 205–206
Anthropic also changed from discarding earlier screenshots to retaining them and compacting the conversation after it exceeds 100,000 tokens. It specifies a 1080p Ubuntu environment, a maximum of 500 actions, max reasoning effort, and five independent runs. The system card cautions against directly comparing the result with earlier task sets or different execution setups.
Partial credit matters. Alongside Opus's 81.8% score, the share of tasks that satisfied every check was 48.7%. The 81.8% figure should not be read as the proportion of jobs fully completed, nor plotted beside Sol's result as if both were directly comparable success rates. System card, page 206
Some practical workloads do have published direct comparisons. On GDP.pdf, which asks expert questions about complex PDFs, OpenAI reports that GPT-6.1 Sol outperformed Opus 5.5 with model fallback across the reasoning settings tested, at less than half the cost per task. On AutomationBench, it says Sol at medium scored 2.2 points above Opus at roughly one-third the cost. OpenAI's practical-work evaluations
These are OpenAI's published comparisons, not LATENT measurements. The announcement text does not establish whether AutomationBench's 2.2-point difference is statistically meaningful. The Opus score in Anthropic's announcement was measured by Zapier during early access, without model fallback and with safety-triggered refusals counted as failures.
| Work evaluated | Published result reviewed | Limits to the comparison |
|---|---|---|
| Understanding PDFs | OpenAI reports Sol outperforming Opus | Tests questions about PDFs, not the entire process of producing a deliverable |
| Multi-step business workflows | OpenAI reports a score and cost advantage for Sol medium | Requires matching benchmark versions, reasoning settings, and fallback conditions |
| On-screen actions | Both companies report gains over their previous models | Task-set dates and history management differ; a direct ranking is not justified |
Extracting numbers from a document to answer a question and using those numbers to build a spreadsheet, enter data into an app and finish the job involve different work and different points of failure. We should consider that distinction when choosing for our own use. Reading scores alone do not establish that a model can be trusted with operations through to delivery.
The API price gap narrows on long inputs
The standard price difference is clear. The table below lists standard direct-API rates in US dollars per million tokens. Tokens are small processing units and do not correspond one-to-one with Japanese characters. These rates are also separate from monthly subscription prices and usage allowances. OpenAI pricing and specifications; Anthropic pricing
| Billing category | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|
| Standard input | $2 | $4 |
| Cache reads | $0.10 | $0.20 |
| Cache writes | $2.50 | $5 for 5-minute retention / $8 for 1-hour retention |
| Output | $10 | $20 |
These Sol rates apply to inputs of 272,000 tokens or fewer. Cache rates apply to the portion successfully reused. Faster processing, regional requirements, and tool fees have separate conditions.
Caching reuses computation for material or conversation prefixes supplied repeatedly. Extended development work often involves the same code and instructions, making cheaper cache reads useful. But the initial cache write also costs money. OpenAI's write rate is a separate rate applied to the relevant input portion, rather than a surcharge added to the normal input rate. OpenAI's caching guide
Retention conditions differ too. GPT-5.6 and later models, including GPT-6.1 Sol, have a minimum retention setting of 30 minutes from the last write or reuse. Anthropic offers 5-minute and 1-hour cache writes. Compare not just read prices but whether a cache will remain usable at your actual request intervals. OpenAI retention conditions; Anthropic cache pricing
Long inputs narrow the price gap. Once Sol's input exceeds 272,000 tokens, input and cache rates double and output rates rise by 50% for the entire request. That means $4 for standard input, $0.20 for cache reads, $5 for writes, and $15 for output. Opus 5.5, by contrast, applies its standard rates across its one-million-token context window. Sol's long-context conditions; Claude long-context pricing
Assuming identical billable token counts, a single uncached request would cost the following.
| Assumed usage | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|
| 100,000 input + 10,000 output tokens | $0.30 | $0.60 |
| 300,000 input + 10,000 output tokens | $1.35 | $1.40 |
Rate-based estimates by LATENT. Output means all billable output tokens, including reasoning. The figures exclude cache writes and reads, tools, faster processing, regional requirements, and tax. Tokenization and consumption differ by model, so these are not measured costs for completing the same work.
In the second example, the difference is only five cents. Sending a large collection of documents in full every time produces a different cost advantage from extracting and reusing only what is needed.
Sol supports a 1.05-million-token context window and Opus one million; both have a maximum output of 128,000 tokens. Context is working space for input and generation, not a guarantee that every part of a document filling that space will be read with equal accuracy. Sol specifications; Claude specifications
Distinguish response speed from time to finish the work
Both models provide controls for how much reasoning to use on answers and tasks. Sol supports low, medium, high, xhigh and max, with medium as the default. Opus 5.5 also defaults to medium and offers settings from low to max. Sol does not support none or minimal, and thinking cannot be turned off in Opus 5.5 either. Even when the companies use the same setting names, these are not a common standard for computing effort or waiting time. Sol reasoning settings; Claude's effort guide; model overview
Less reasoning can make an individual attempt lighter. But if a difficult task then needs more trial and error, total completion time may increase. Maximum reasoning does not guarantee a higher score either. On Anthropic's FrontierCode v1.1 results, Opus medium scored 54.6%, compared with 54.4% for max in the performance table. Small differences like these should not be treated as conclusive evidence that one setting is better. Opus results by reasoning effort
Speed claims also need the right comparison. Anthropic's claim of more than 30% faster output compares Opus 5.5 with Opus 5, not Sol. It advertises up to 2.5x speed in a separately priced Fast mode. OpenAI announced that Sol's Ultrafast mode would offer up to 8x token generation speed relative to standard Codex speed in the following days. An announced future multiplier and current standard speed cannot establish a winner. The companies' announcements; Opus speed details
API prices checked September 30 list Sol Fast at $4 for input and $20 for output, and Opus Fast at $8 and $40. Opus Fast is a research preview with restrictions on where it is offered. This comparison also needs to be kept separate from standard pricing. OpenAI pricing; Anthropic pricing
Our document review did not establish a direct comparison of time to first response and time to completion with matching input lengths, reasoning settings, and execution environments. The speed users experience includes waiting for the first response, tool calls, and recovery from failures, as well as the rate at which text appears.
Availability and considerations when delegating actions to AI
At launch, GPT-6.1 Sol was announced for Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, and for the API. OpenAI explicitly said it was not yet available in ordinary Chat. Actual access also depends on the plan, client, and organization settings. Its API identifier is gpt-6.1-sol, and tool-using workflows use the Responses API. OpenAI availability details; model specifications; changelog
Opus 5.5 is announced for Claude, Claude Code, and Claude Platform, as well as AWS, Google Cloud, and Microsoft Azure. Its direct-API identifier is claude-opus-5-5. Pricing and features through cloud providers should not be assumed to match the direct-API rates in this article. Anthropic availability details; models by platform
When delegating actions, the conditions under which a model stops are part of its performance. In Anthropic's evaluations, safety mechanisms can route cyber tasks to Opus 4.8 and tasks such as biology to Opus 5. Results that include fallback do not mean Opus 5.5 alone solved every task. They need to be distinguished from tests without fallback. Opus 5.5 evaluation conditions
OpenAI likewise says it applies Astra's safety measures to GPT-6.1 Sol. It evaluates attempts to bypass automated-review refusals and to perform actions the user did not authorize. These tests deliberately collect difficult situations; they do not measure the incident rate in ordinary use. GPT-6.1 Sol system card
Whichever model you choose, better instruction-following does not mean every action can be delegated without review. Deciding where a person checks difficult-to-reverse actions, such as sending email or deploying to production, belongs alongside model selection.
Choose for your own workload
GPT-6.1 Sol is worth comparing if you are considering Opus 5.5. Its lower standard and cache rates could be particularly useful for repeated code changes or document processing that reuse the same background material. Public evaluations also show improvements in practical capabilities that support that proposition.
The case for switching on price alone is weaker when long inputs narrow the gap, or when Opus finishes a task with fewer retries. This document comparison cannot establish which workloads fall into the latter category.
One approach is to select examples from everyday work, such as fixing existing code, preparing a report from long documents and completing a routine task involving screen interaction, then compare both models against the same acceptance criteria for each. Record the billing details, time to completion, corrections made by a person and reasons for stopping. Separating work that ordinary reasoning can handle from work that benefits from more reasoning removes the need to use the highest-scoring model every time.
This release does not settle a universal winner or loser against Opus 5.5. It adds a lower-cost option worth comparing when delegating complex work. The place to test that difference is the process of finishing your own tasks, not a single line in an overall ranking.