Publications

    Intelligence artificielle

    GPT-6.1 Sol on Vercel AI Gateway: announced costs and relevance for SMBs

    Vercel announces GPT-6.1 Sol in AI Gateway. Its published rates are lower than Astra’s, but Vercel’s billed costs and actual gains for SMBs still need to be validated. Here are the available figures and the criteria for a useful trial.

    Deux dossiers identiques examinés à la loupe; à l’arrière, un cube fait face à cinq cubes de même taille.

    Vercel announces that GPT-6.1 Sol, presented as an OpenAI model, is accessible in Vercel AI Gateway. Highlighted uses include debugging, using computer applications, multi-step business processes and finding answers in PDFs containing tables and charts.

    For a Quebec SMB, the potential value lies in evaluating this model on demanding tasks. However, the announcement demonstrates neither a productivity gain for any given business nor a reduction in the total cost of an automation.

    Source limitation: this article is based solely on the excerpts provided. The pages, APIs and actual availability of the model have not been verified live. Performance results are those reported by the providers.

    Announced availability, to be confirmed for your account

    According to Vercel’s announcement, the identifier to use is openai/gpt-6.1-sol. Vercel documents its use with the AI SDK, the OpenAI-compatible Chat Completions API and the Responses API.

    These details describe an announced integration, not one tested for this article. Account-specific access, quotas and rate limits remain to be confirmed. The excerpts also do not establish the exact date API access became available.

    Standard rates at one-fifth of Astra’s

    The excerpt published by OpenAI presents the following rates per million tokens. The currency is not specified in the excerpt: the “$” symbol is reproduced as given.

    | Model | Standard input | Output | Cached input | |---|---:|---:|---:| | GPT-6.1 Sol | $2 | $10 | $0.10 | | GPT-6 Astra | $10 | $50 | $1 | | GPT-6 Luna | $0.10 | $0.50 | $0.01 |

    GPT-6.1 Sol’s standard input and output rates therefore represent 20% of Astra’s, an 80% reduction. This does not mean a business process will automatically cost 80% less: token consumption, retries and human review must also be evaluated.

    Another important distinction: these amounts are OpenAI’s published rates, not confirmation of the final amount billed by Vercel. Any additional fees are not established by the excerpts.

    Caching: conditional savings

    At $0.10 per million tokens, cached input is priced 95% lower than Sol 6.1’s standard input. Requests that reuse a long shared context, such as internal procedures, could therefore cost less.

    However, this possibility depends on the caching terms, request eligibility and actual cache use. The discount does not apply to output rates and does not guarantee an equivalent reduction in the total bill.

    Reported improvements, not proof of performance for SMBs

    OpenAI positions GPT-6.1 Sol as an upgrade to GPT-6 Sol that approaches Astra in certain evaluations, at a lower cost. This does not amount to general equivalence between the models.

    The provider reports, in particular:

    • DeepSWE v1.1: a gain of 6.4 percentage points over GPT-6 Sol’s best score, with lower reasoning effort and cost;
    • AutomationBench: a gain of 4.8 points over GPT-6 Sol at the same medium reasoning setting;
    • OSWorld 2.0, offline set: a gain of 7 points over GPT-6 Sol at maximum reasoning effort, at less than half the cost.

    The excerpts do not provide all absolute scores or the complete protocol for reproducing these comparisons. These results are indicators to examine, not a measure of reliability on an SMB’s files or in Quebec French.

    Three targeted trials before migrating

    Based on the announced uses, three trials would be relevant:

    1. Software development: fix a known bug, then verify the fix with tests.
    2. Document analysis: answer questions about complex PDFs, comparing the answers with the source passages, tables or charts.
    3. Administrative automation: execute a multi-step process in a test environment, with human approval for sensitive actions.

    The comparison should use the same tasks with the current model and Sol 6.1. The metrics to track are cost per accepted task, error rate, latency and human review time. Acceptance criteria should be defined before the trial.

    A lower price per token is not enough to demonstrate better overall performance.

    Checks to complete before production

    The excerpts leave several questions open: context window, file limits, caching terms, costs billed by Vercel and quality on the business’s own documents.

    Data residency, retention and use must also be clarified, as must compliance with Quebec’s personal information protection requirements. The announcement does not establish such compliance.

    In brief: GPT-6.1 Sol is an announced option to evaluate for demanding automations and reused contexts. The decision should rest on a representative trial and validation of the technical and contractual terms, rather than on rates or reported results alone.

    Have a similar project?

    Tell us what you need: we reply with a clear first read, no strings attached.