Sign in with ChatGPT meets GPT-6.1 Sol: the allowance is the story

What does sign in with ChatGPT actually unlock?#

Sign in with ChatGPT is the headline, but the allowance is the harder detail. OpenAI used its 29 September developer event to make a wider promise: OpenAI’s DevDay recap says a user can carry an existing ChatGPT allowance into eligible work across 16 partners, including Cognition’s Devin, Notion and Vercel.

The same announcement says the user can control how much each partner consumes. That is a useful opening for people who have paid for one plan and found a second meter waiting inside every other tool. It is also narrower than an unlimited pass. Eligible OpenAI usage counts towards the plan limits; the partner may still have its own paid features, terms and availability. The GPT-6.1 Sol launch makes that route more consequential because the model now sits where many of those tools do their work.

There are two doors in the den. The first is identity: OpenAI’s sign-in help page says ordinary sign-in shares a name, email address and profile picture, if present. It does not independently hand over ChatGPT conversations, files, tokens or billing information. The second is model usage.

OpenAI’s plan-usage documentation calls that an optional capability which asks for permission to use a user’s plan for eligible Responses API requests. A neat sign-in button does not, on its own, establish that a product can spend the allowance. Readers should watch for the separate consent screen and for the partner’s own explanation of limits. Sign in with ChatGPT should be treated as the start of a permission review, not the end of one.

OpenAI’s identity help page names Airtable, GitLab, HubSpot, Notion, Supabase and Vercel among initial sign-in partners. Its DevDay account names a broader plan-usage set. Those lists answer different questions. A partner can support convenient account creation without offering plan-backed model calls, and plan-backed access can be limited to particular clients or workflows. The AI consultancy buyer should ask which path is being offered, which model it reaches and where the meter is visible before moving a real team into it. ChatGPT plan usage and simple identity sign-in need separate scrutiny.

For a subscriber, Sign in with ChatGPT needs a visible account path, a visible allowance and a visible stop control. Kilo provides a helpful concrete case. Kilo’s provider documentation says its supported local coding surfaces can authenticate with a Plus or Pro subscription, show the relevant Codex usage windows and use the subscription allowance without a separate API charge.

Kilo also says cloud features are excluded from that route. That distinction matters more than the bright button: a developer may save money in a local editor while a cloud agent remains on a different bill. The same pattern should be checked in every partner product rather than assumed from a logo wall. Sign in with ChatGPT in Kilo therefore answers a more precise question than a generic promise of access.

Where is GPT-6.1 Sol available on launch day?#

OpenAI’s GPT-6.1 Sol announcement says the new model is available on 29 September in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and through the API as gpt-6.1-sol. It says the model is not yet in ordinary Chat. That surface split is easy to miss when a headline simply says “available”.

A person who opens a normal ChatGPT conversation should not assume the new model is on that screen. A team assessing a partner integration should check the model list inside the partner itself. This OpenAI model release is aimed at professional work, where access through an existing plan may matter as much as a benchmark score. The Sign in with ChatGPT rollout changes how some users reach these surfaces, subject to each partner’s support.

The model is the second half of the story because subscription portability changes the route to it. OpenAI positions GPT-6.1 Sol as an upgrade for coding, computer use and professional tasks. the GPT-6.1 Sol model page sets out the API model identifier and capabilities, while the announcement gives launch-day availability.

The practical trail is therefore: verify the partner supports plan usage, verify the client exposes the model you want, then test that the job completes within the allowance. One can fail while the other two look fine. Sign in with ChatGPT is a route to check, not proof that any one partner has switched on this exact model.

The published standard API price is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million. OpenAI’s API price list is the place to verify the current rate and its context tiers. OpenAI says this makes standard input and output token rates one fifth of Astra’s.

It does not mean every real task costs one fifth as much: prompts, reasoning effort, retries, tool calls and cache hit rates all move the bill. Nor does a plan-backed partner make an API invoice disappear for a separate API key. Work through a representative task, then compare completed work per pound rather than the headline rate per token. AI model pricing is useful evidence, but the relevant bill is the one for a completed task. Sign in with ChatGPT may be the cheaper route for an eligible subscriber, but only within the documented allowance.

That price chart is a rate card, not a cost-per-outcome chart. A sharp shopping list for a team includes latency, model availability, the partner’s service charge and the length of its work. The published folkfox rates are a reminder that a useful unit is the complete job and the human review it still needs. GPT-6.1 Sol may change the choice of default model. It does not remove the need to measure the complete workflow.

What do the GPT-6.1 benchmarks show, and who measured them?#

The strongest current numbers are launch figures reported by OpenAI. On DeepSWE 1.1, the company says GPT-6.1 Sol matches Astra at roughly a fifth of the cost and beats the best GPT-6 Sol score by 6.4 percentage points. The the DeepSWE leaderboard explains its 113 long-horizon coding tasks and verification method, but its public leaderboard was last updated on 22 September when checked for this article, before GPT-6.1 Sol launched.

It therefore cannot yet independently confirm the new model’s entry. That is a material limit, not a small-print footnote. For AI coding models, that independent leaderboard gap is especially important because the launch claim is about coding work. A Sign in with ChatGPT connection cannot settle that evidential gap.

OpenAI also says GPT-6.1 Sol beats Opus 5.5 with fallbacks on the GDP.pdf document benchmark at less than half the tested cost per task. Surge AI’s GDP.pdf benchmark describes a demanding set of professional PDFs drawn from finance, healthcare, legal and other work. Its public leaderboard likewise did not show a new GPT-6.1 Sol result at our check.

Across another business-workflow measure, OpenAI reports a 2.2 percentage-point lead over Opus 5.5 at medium effort and a 4.8-point gain over GPT-6 Sol. Zapier’s AutomationBench describes AutomationBench as end-to-end tests with 47 tools, scored on the final state of the work. These are useful workload shapes, but the new model comparisons in this article remain the vendor’s reported evaluations. Sign in with ChatGPT widens access to a model; it does not change who ran the test.

Computer use is another promising trail. OpenAI reports a seven-point improvement over GPT-6 Sol on the OSWorld 2.0 offline set at maximum effort, coming within 2.1 points of Astra at roughly one seventh of the cost per task. the OSWorld 2.0 benchmark describes what its long-horizon desktop tasks are meant to test.

Scientific work tells a more mixed tale: OpenAI says GPT-6.1 Sol more than doubles GPT-6 Sol’s score on Terminal-Bench Science 0.1 at maximum effort, yet Astra still leads the tested models at 68.1 per cent. Terminal-Bench Science gives the benchmark context. That means “near Astra” is a workload-dependent claim, not a blanket ranking.

The most legible numerical comparison is factuality, with an unusually important caveat. OpenAI says that on a set of conversations chosen because users had flagged earlier factual errors, the share of low-effort answers with at least one error fell from 11.4 per cent for GPT-6 Sol to 7.7 per cent for GPT-6.1 Sol. This is a 3.7-point absolute fall and about a 32 per cent relative reduction. The company explicitly says these hard prompts are not representative of normal use. Presenting 7.7 per cent as a general hallucination rate would mislead readers.

OpenAI’s system-card addendum says its alignment checks deliberately stress difficult situations. It reports improvement in respecting restrictions, disclosing a broken search tool and avoiding unauthorised outcomes compared with GPT-6 Sol. The tests do not tell a buyer how often a normal task will go wrong in production. A calm comparison keeps the three layers separate: OpenAI’s lab result, an independent benchmark once it actually publishes the new model, and your own task completion rate. The fox should follow the strongest scent, but still inspect the ground. A credible AI model evaluation should record those three layers before naming a winner.

How should a team test the subscription route?#

Start with a bounded pilot: a dozen ordinary jobs and two or three difficult ones from your own workflow. For coding, include a failing test, a change across files and a reviewable explanation. For professional documents, include a dense table and an answer that must cite the right cell. For browser work, include a task where the agent must stop at a permission boundary. The launch-day GPT-6.1 benchmarks are leads for where to test first, not promises that your jobs will land at the same score. Sign in with ChatGPT belongs in the pilot only when its exact permission and usage path is known.

Put the allowance in the pilot sheet. The partner should show which account is connected, which model is used, what action consumes the plan, and how a user disconnects it. OpenAI says eligible partner use counts towards plan limits; Kilo documents its own visible usage windows. A quiet quota is better than a surprise invoice, but a hidden quota can still stop work at the wrong hour.

Ask whether partner seats, hosting and storage are charged separately. The brand strategy question is simple: does the product’s promise match the purchase path the customer actually meets? ChatGPT subscription partners may each draw a different boundary around the eligible work.

For a team in a regulated category, permissions and provenance deserve the same care as cost. Review what identity data goes to the partner, what model access is separately authorised, which workspace policy applies and what audit trail remains when an agent acts. OpenAI’s sign-in help page says administrators can control identity sign-in for organisations, while delegated access can need separate approval. The cybersecurity work lesson is to treat the approval screen as a product surface, not a legal afterthought. Sign in with ChatGPT also needs a clear exit route when that boundary changes.

Finally, compare the outcome with a baseline. If GPT-6.1 Sol finishes more jobs, costs less per accepted result or needs fewer human corrections, there is a real case to change. If the partner’s supported surface is a narrow editor while the team’s work lives in a cloud agent, plan sharing may not cover the work that matters. If the best result still needs Astra, say so.

The content team and app team should measure their own review minutes as carefully as token prices. A cleaner sign-in is welcome. A cleaner decision needs receipts. Sign in with ChatGPT should lower friction without lowering the standard of proof.

• 16

  partners named by OpenAI for plan use

• 2

  separate consent questions: identity and plan

• 0

  independent GPT-6.1 Sol public leaderboard entries verified at publication

The final move is editorial as well as technical. Do not call the sign-in route a free API, and do not call a vendor benchmark an independent result. In our interactive story on claims under uncertainty, the rule is that a headline should never be firmer than its evidence. Here the evidence is clear enough for a useful buying test: eligible subscription use is expanding, GPT-6.1 Sol is available in Work and Codex, and the strongest performance figures still need independent and local verification.

Frequently asked questions#