On September 22, 2026, OpenAI announced GPT-6.1 Sol, an upgraded version of its mid-tier AI model. It’s available through the OpenAI API and is rolling out inside ChatGPT Work and Codex for eligible Plus, Pro, Business, Enterprise, and Edu users. On October 22, The New Stack reported that OpenAI’s internal COBOL migration test reached 86.6% accuracy at low reasoning effort and 97.9% at higher effort. That means turning decades-old business code into modern programming languages. (For the launch basics, see our GPT-6.1 Sol availability and pricing explainer.)
So should you care? If you run a bank, an insurer, or a government IT shop, this is worth watching. If you’re an everyday ChatGPT user, probably not yet, because OpenAI says the model isn’t yet available in regular ChatGPT chats. The eye-popping COBOL numbers also come from OpenAI itself, and nobody outside the company can check them yet. (Sources below were checked October 24, 2026.)
What Happened
OpenAI’s official announcement describes GPT-6.1 Sol as an update to its “Sol” model. Sol sits in the middle of OpenAI’s current lineup:
- GPT-6 Luna: The fast, cheap model for everyday work at scale.
- GPT-6.1 Sol: The middle option. It replaces the earlier GPT-6 Sol.
- GPT-6 Astra: OpenAI’s flagship and, by the company’s own description, its most intelligent model.
According to OpenAI’s launch announcement and model documentation, the lineup is positioned around three roles: choose GPT-6.1 Sol “to balance intelligence and cost,” use Luna for cost-sensitive, high-volume work, and reserve Astra as the flagship for the most demanding reasoning and coding tasks.
If those names feel like a box of cereal brands, you’re not alone. Think of it like coffee sizes. Luna is the small, Astra is the extra-large, and Sol is the medium that OpenAI now says tastes almost as good as the extra-large.
OpenAI’s main pitch in its launch announcement (Introducing GPT-6.1 Sol) is that GPT-6.1 Sol approaches Astra on selected agentic coding, computer-use, and professional-work evaluations (such as DeepSWE v1.1, GDP.pdf, AutomationBench, and OSWorld 2.0). In an agent task, the AI works through a chain of steps on its own. It doesn’t stop after answering one question. OpenAI’s GPT-6.1 Sol model documentation sums the model up as delivering “near-Astra performance for complex work at a lower cost.”
Some other confirmed details come from OpenAI’s API model documentation and models page:
- Model ID:
gpt-6.1-solfor developers using the API. - Reasoning effort levels: The model offers five reasoning effort settings: low, medium (the default), high, xhigh, and max. They control how much the model “thinks” before it answers. More thinking usually means better accuracy. It also costs more and takes longer.
- Context window: In the API, GPT-6.1 Sol has a 1,050,000-token context window (up to 922,000 input tokens and 128,000 output tokens in a single response), per OpenAI’s GPT-6.1 Sol model documentation. A token is a piece of text as the model’s tokenizer splits it; in English it’s often part of a word, but code can split differently. Product interfaces such as ChatGPT Work and Codex have their own product-level limits.
- Accuracy and safety: OpenAI describes GPT-6.1 Sol as an upgrade over the previous GPT-6 Sol. Its deployment safety documentation covers the safety testing separately.
The COBOL Claim, in Plain English
This is the part that made headlines.
COBOL is a programming language from the late 1950s. That sounds like ancient history, but as The New Stack notes in its October 22, 2026 report, legacy COBOL programs are still used in banking, insurance, and government systems. To illustrate why enterprises care, The New Stack cites older industry estimates originally attributed to IBM claiming COBOL underpins over 40% of online banking systems, supports 80% of in-person credit card transactions, and handles 95% of ATM transactions (though neither the original survey date nor geographic scope is specified in the report). The New Stack also points to two reasons legacy modernization projects stall in practice: shrinking pools of engineers familiar with older mainframe codebases and documentation that has fallen behind decades of patches. That combination makes full system replacements feel a bit like swapping the foundation of a house while people still live in it.
According to OpenAI’s internal COBOL migration evaluation (published October 19, 2026 and covered in The New Stack’s October 22 report, “OpenAI’s COBOL migration score has made a striking leap”), the benchmark evaluates migrating legacy COBOL programs into modern target languages while preserving observable program behavior across hidden test suites. Because the evaluation is internal to OpenAI and has not been independently reproduced, the headline figures below reflect OpenAI’s reported benchmark scores:
| Model (as reported) | Reasoning effort | COBOL migration accuracy |
|---|---|---|
| GPT-5.6 Sol (as labeled in the report) | Not specified | 5.4% |
| GPT-6.1 Sol | Low | 86.6% |
| GPT-6.1 Sol | Higher setting | 97.9% |
Source: OpenAI’s internal evaluation, as reported by The New Stack. Nobody outside OpenAI has reproduced these figures.
The New Stack reports that OpenAI’s low-effort test run cost 39% less than the prior model’s run under its undisclosed test conditions; because neither OpenAI nor The New Stack publishes the underlying token counts, tool-call volume, or reasoning settings, that figure applies only to that benchmark run rather than general migration costs.
A jump from 5% to nearly 98% is big enough to make you do a double take. That’s exactly why the fine print matters.
The Fine Print You Shouldn’t Skip
- It’s OpenAI grading OpenAI: These numbers come from OpenAI’s own internal test. The company hasn’t published the test cases or dataset. Outside researchers can’t currently rerun it or check the results.
- A naming wrinkle: The New Stack’s report names the baseline model “GPT-5.6 Sol” (“up from GPT-5.6 Sol’s 5.4%”). OpenAI’s launch materials call Sol’s direct predecessor “GPT-6 Sol.” The public reporting doesn’t settle which model produced the 5.4% baseline, so treat that comparison with some caution.
- Humans still required: As quoted by The New Stack, OpenAI says “applying it to production software still requires expert review and testing against the system’s business rules and dependencies.” Our own advice goes further: treat every migrated module as untested until your team proves it behaves exactly like the old one, down to every odd business rule from 1987.
So are COBOL programmers suddenly out of a job? No. Even if you take OpenAI’s numbers at face value, someone has to check the AI’s work. Someone has to test it and sign off before a bank trusts it with real accounts. If anything, people who understand these old systems become the quality inspectors. That job still needs deep expertise.
Who Can Actually Use It
This is where a lot of readers will hit a wall, so here’s the plain version:
- Regular ChatGPT chats: GPT-6.1 Sol is not yet available in regular ChatGPT conversations.
- ChatGPT Work and Codex: It’s rolling out to eligible Plus, Pro, Business, Enterprise, and Edu users in these tools. OpenAI’s Help Center says availability depends on your plan, workspace settings, and rollout access, and Enterprise and Edu admins must enable it.
- Developers: API users can call it as
gpt-6.1-sol, billed per token. Rate limits depend on your API usage tier.
If you open a regular ChatGPT chat and don’t see GPT-6.1 Sol, nothing is broken: it isn’t available there yet. In Work and Codex, whether you see it can also depend on your plan, workspace settings, and rollout access.
How It Stacks Up
Against OpenAI’s Own Lineup
Here’s the published standard API pricing per million tokens for the Sol family, from OpenAI’s model pages for GPT-6.1 Sol and GPT-6 Sol and the API pricing documentation:
| Model | Standard input (≤272K) | Standard output (≤272K) | Standard cached input (≤272K) | Cache writes (30m / 12h) | Long-context standard (>272K input) |
|---|---|---|---|---|---|
GPT-6.1 Sol (gpt-6.1-sol) | $2.00 | $10.00 | $0.10 | $2.50 / $5.00 | $4.00 input / $15.00 output / $0.20 cached |
GPT-6 Sol (gpt-6-sol, prior version) | $2.00 | $10.00 | $0.20 | $2.50 / $5.00 | $4.00 input / $15.00 output / $0.40 cached |
Standard API service-tier rates per million tokens from OpenAI’s model and pricing documentation, checked October 24, 2026. These per-token rates apply to API calls; ChatGPT Work and Codex usage is governed by the applicable subscription plan terms.
Compared with the earlier GPT-6 Sol, standard input ($2.00) and output ($10.00) rates are unchanged, while cached input drops in half from $0.20 to $0.10 per million tokens.
How does GPT-6.1 Sol compare with GPT-6 Astra and GPT-6 Luna?
- Versus GPT-6 Astra: OpenAI’s GPT-6.1 Sol model documentation positions Sol as delivering “near-Astra performance for complex work at a lower cost,” reserving GPT-6 Astra as the higher-priced flagship for the most demanding reasoning and coding workloads.
- Versus GPT-6 Luna: In its launch announcement and API pricing documentation, OpenAI positions GPT-6 Luna as the lower-cost model for fast, high-volume everyday work.
Separate API-only service-mode modifiers also apply on top of the standard schedule above:
- Long-context requests (>272K input tokens): Billed at 2x standard input and cached-input rates and 1.5x standard output rates for the entire request ($4.00 input, $0.20 cached input, and $15.00 output per million tokens on GPT-6.1 Sol).
- Batch and Flex processing: Billed at 50% below Standard rates.
- Fast mode: Billed at 2x Standard rates.
- Ultrafast mode: Billed at 6x Standard rates (via
service_tier: "ultrafast"in the Responses API; note that 6x is the API billing multiplier, distinct from OpenAI’s separate launch claim that Ultrafast delivers up to 8x faster token generation in Codex). - Cache writes: Billed at $2.50 per million tokens for 30-minute retention (1.25x standard input) and $5.00 per million tokens for 12-hour retention (2.5x standard input).
- Regional processing (data residency): Adds a 10% uplift where available.
That cached input price matters for big legacy-code projects. “Cached” means the model reuses context it has already seen, like the same large codebase across many requests. For GPT-6.1 Sol, cached input is 95% cheaper than fresh input ($0.10 versus $2.00), while writing to the cache costs $2.50 per million tokens for 30-minute retention and $5.00 per million tokens for 12-hour retention. Because lines of COBOL don’t map neatly to tokens, large migrations usually work in smaller modules rather than feeding an entire repository into one prompt.
Against Anthropic’s Claude Opus 5.5
In its launch announcement, OpenAI says GPT-6.1 Sol scored higher than Anthropic’s Claude Opus 5.5 at lower task cost on two specific evaluations cited by OpenAI (GDP.pdf and AutomationBench), without publishing full cross-vendor methodology or per-task cost tables. As separate context, The New Stack’s October 22 report notes that Real-SWE 1.1—an independent software-engineering benchmark built on private codebases—listed Claude Opus 5.5 at 56.8% task resolution; because The New Stack cites only that isolated figure without a full leaderboard or methodology, and because Real-SWE 1.1 does not test COBOL translation, it shouldn’t be read as a direct head-to-head ranking for legacy migration.
The OpenAI launch materials cited here do not include a benchmark comparison with a current Google Gemini model, so we won’t guess at one.
The Reaction
Public reaction is still forming. The New Stack emphasized that developers “cannot independently inspect” the COBOL evaluation. It also reports that OpenAI released the setup and test suite for its separate MultiLangBench benchmark on GitHub.
Our Take
Verdict: Real news for legacy-heavy organizations, mostly background noise for everyone else, and the COBOL numbers need an outside referee.
Here’s who should actually pay attention:
- Banks, insurers, government agencies, and their IT contractors: Yes, get curious. If the reported COBOL results hold up, this could change the math on modernization projects. For organizations that have postponed modernization because of cost or risk, the reported results may justify a small pilot. Just don’t sign a big contract based on one vendor’s internal test. Run your own pilot on a small, low-risk piece of code first.
- Developers already paying for Astra: This is the most practical win. With GPT-6.1 Sol priced at $2.00 input, $10.00 output, and $0.10 cached input per million tokens in the Standard tier, OpenAI positions it as a lower-cost alternative to flagship Astra workloads. Testing it on your own modules will tell you quickly whether its quality is close enough to Astra to switch.
- Small business owners: Interesting, but unlikely to change your week. Unless you run custom legacy software, nothing here affects the tools you use every day.
- Everyday ChatGPT users: You can skip this one. It is not available in regular ChatGPT chats, so most everyday users do not need to change plans just to use it.
Our honest opinion: jumping from roughly 5% to nearly 98% on any code-migration test should make engineering leaders curious, and OpenAI deserves credit for reminding teams that human review against business rules is still mandatory before touching production.
Wrapping Up
OpenAI positions GPT-6.1 Sol as a lower-cost middle-tier model for coding and multi-step agent workflows. If you work somewhere that still runs on decades-old mainframe code, it’s worth a small pilot on non-critical modules. If you just use ChatGPT to plan dinners and draft emails, you can safely let this one pass you by.
| What | Details | Applies To |
|---|---|---|
| Announced | September 22, 2026 | Everyone |
| COBOL claim | 86.6% (low effort) to 97.9% (higher effort) on OpenAI’s own behavior-parity test, as reported by The New Stack; not independently reproduced | Legacy-system owners |
| Price | $2 input / $10 output per million tokens (standard, up to 272K input tokens) | API developers |
| Access | API, plus a rollout in ChatGPT Work and Codex for eligible Plus, Pro, Business, Enterprise, and Edu users; not yet in regular ChatGPT chats | Paid plans, developers |
| Next step | Pilot it on a small, low-risk code sample before trusting the headline numbers | Businesses with legacy code |