GPT-6.1 Sol: 5 Honest Price Checks Before You Switch

OpenAI released GPT-6.1 Sol at DevDay on September 29, 2026, and the number that decides everything else is $2 per million input tokens against GPT-6 Astra’s $10. I pay for ChatGPT, Claude and Cursor every month, and I have not run GPT-6.1 Sol on a single client file — it shipped yesterday. So this is a read of the published numbers, not a verdict from my own invoices. What I can do is hold the price sheet against the benchmark table, which is where most model launches quietly come apart, and work out which parts of a one-person stack the arithmetic actually moves.

GPT-6.1 Sol Undercuts Astra by 80% on the Published Rate Card

The headline is a straight 80% cut on both sides of the ledger. GPT-6.1 Sol lists at $2.00 per million input tokens and $10.00 per million output tokens, where GPT-6 Astra lists at $10.00 and $50.00. That is the entire substance of OpenAI’s claim that the model delivers near-Astra intelligence “at one-fifth of Astra’s standard API input and output token prices,” and unlike most launch-day framing it survives contact with the pricing page.

Cached input is where the GPT-6.1 Sol gap widens further. GPT-6.1 Sol reads cached input at $0.10 per million — 95% below its own standard input rate, half of what GPT-6 Sol charged, and a tenth of Astra’s $1.00. For any workflow carrying a large stable system prompt, that one number moves more money than the headline rate does.

Two footnotes matter before anyone builds a budget on this:

Long context reprices the whole request. Once a prompt crosses 272,000 input tokens, the entire call bills at $4.00 input, $0.20 cached input and $15.00 output per million. Not just the overflow — all of it.
Batch halves it again. Asynchronous batch requests run at $1.00 input and $5.00 output per million, with cache reads at $0.05.

That long-context cliff is the part I would expect people to get wrong first. A 300,000-token repository dump does not cost you a small surcharge on the last 28,000 tokens. It doubles the input rate on everything you sent.

The Benchmark Gap Is One to Two Points, Not a Tier Drop

A price cut of this size normally means you are buying a smaller model, and here the published evals say otherwise. On DeepSWE v1.1, GPT-6.1 Sol scores 75.2 against Astra’s 74.8 and GPT-6 Sol’s 68.8 — ahead of the flagship on OpenAI’s own software-engineering eval. On OSWorld 2.0 computer use it lands at 71.4 against Astra’s 73.5, and on the GDP.pdf document benchmark it posts 32.0 against 32.2.

So the pattern is one win, one two-point loss, one statistical tie. Set against the cost-per-task figures OpenAI published alongside them, that is an unusual trade:

DeepSWE v1.1: roughly $1.50 per task on GPT-6.1 Sol against about $7.70 on Astra
OSWorld 2.0: about $1.30 per completed task against roughly $9.30 — one-seventh the cost

Both models carry the same 1,050,000-token context window and the same 128,000-token maximum output, so this is not a capability tier being quietly trimmed to hit a price. GPT-6.1 Sol ships the same envelope at a fifth of the rate, with a benchmark spread narrow enough that on most work it would take a careful A/B to notice.

The honest caveat is that all of these numbers are the vendor’s. Nobody outside OpenAI has had GPT-6.1 Sol long enough to publish an independent replication, and a two-point OSWorld gap is exactly the kind of margin that widens on messier real tasks than a benchmark harness produces. When Astra landed I ran through a similar set of checks before letting it near billable work, and the launch table looked cleaner then than the first month did.

GPT-6.1 Sol Reprices Agent Loops, Not Chat Windows

If you mostly type questions into a chat box, this launch changes nothing about your month. The Plus subscription you already pay for does not bill per token, and an 80% API cut is invisible from inside it. The people who feel this are the ones running loops — Codex tasks, agent runs, anything where a model calls tools eleven times to produce four lines of visible output.

That is where the arithmetic gets interesting, because agent loops are dominated by the invisible half of the bill. I have written before about how much of an agent’s output spend is deliberation rather than answer, and the short version is that the runs which look cheapest on screen are frequently the most expensive on the invoice. Cutting the output rate from $50 to $10 per million hits that hidden portion at full strength.

A cheaper model does not save a solo operator money on the work they can see. It saves money on the work the agent did while they were looking somewhere else.

Stacked with the $0.10 cached input rate, the realistic saving on a long-running agent task is larger than the 80% headline, because caching and output pricing both improved at once. OpenAI’s own cost-per-task figures — a seventh of Astra on OSWorld — are the clearest evidence of that compounding.

There is a second consequence worth naming. At $2 input, work that previously did not justify a frontier model starts to. Bulk classification, first-pass document triage, the sort of job I currently hand to a smaller model and then re-check: the gap between “good enough model” and “best available model” narrows to a rounding error on those tasks. That is a routing change more than a switching decision.

Ultrafast Costs 6x and Has Not Shipped for Sol Yet

The speed tier is the part of this announcement that is not yet real. OpenAI says a GPT-6.1 Sol Ultrafast option will deliver up to eight times the standard token generation rate in Codex — reported at up to 300 tokens per second — at six times standard pricing, arriving “in the coming days.” Ultrafast is live for Astra today; for Sol it is a promise with a date attached to it loosely.

Run that multiplier before treating it as a cheap tier. Six times $2 and $10 is $12 input and $60 output per million, which is above Astra’s standard rate on both sides. Ultrafast is not GPT-6.1 Sol going faster at a discount; it is a latency product priced accordingly, and the moment you enable it the entire cost argument for GPT-6.1 Sol inverts.

The same DevDay brought a new $500 ChatGPT Pro tier bundling Ultrafast access and 25x the Plus usage quota, plus dots — always-on agents on Pro and Business Premium, each running on Astra with its own cloud computer and browser. For a one-person consultancy, both of those are the wrong end of the announcement. The $2 rate card is the part that reaches a solo budget.

What I Would Check Before Routing Paid Work to GPT-6.1 Sol

Nothing here needs a decision this week, and launch week is the worst possible time to move a client workflow. These are the five checks I would run first, in order:

Measure your current agent spend before comparing rates. If you have never opened the reasoning-token breakdown on your existing provider, an 80% cut on a number you cannot see is not yet a saving you can bank.
Check whether your prompts cross 272,000 tokens. Long-context requests reprice at $4 and $15 per million, which is still cheaper than Astra but destroys the tidy fifth-of-the-price arithmetic.
Confirm the cache actually hits. The $0.10 cached input rate carries most of the benefit on repeated runs, and a system prompt that changes between calls never gets there.
Re-run your own worst task, not the benchmark. A two-point OSWorld gap is small; whether it shows up on your specific messy input is the only test that counts.
Leave Ultrafast switched off until you have priced it. At 6x, it costs more than the flagship it was meant to undercut.

Against my own stack, the comparison I would want is not Sol against Astra but Sol against what I already pay for. Anthropic moved its own pricing recently and I wrote up what a lower Opus rate did and did not change for daily work; the conclusion there was that a cheaper model only helps if it replaces something, and a second subscription running beside the first is a cost increase wearing a discount’s clothing.

For me, GPT-6.1 Sol is a routing candidate rather than a migration. I will keep Claude and Claude Code where they are, because the drafting and file-automation work they do is not priced per token in my setup and switching cost is real. What I will do is open the reasoning-token breakdown on the agentic jobs I run through Codex-shaped tooling and see how much of that spend a $2 input rate would actually reclaim. If the answer is meaningful, the model can earn a narrow lane; if it is noise, the cheapest frontier model of the week is still just a headline.

Sources

AI-assisted research and drafting. Reviewed and published by ToolMint.

In this article

GPT-6.1 Sol Undercuts Astra by 80% on the Published Rate Card

The Benchmark Gap Is One to Two Points, Not a Tier Drop

GPT-6.1 Sol Reprices Agent Loops, Not Chat Windows

Ultrafast Costs 6x and Has Not Shipped for Sol Yet

What I Would Check Before Routing Paid Work to GPT-6.1 Sol

Sources