Coding Horizon

GPT 6 sol’s 91% cost saving puts claude on notice

Every figure, price, date and benchmark score this video states, chased to a primary source. Where a number is somebody’s own reported result rather than an independent measurement, that is said here and the reporter is named, because several of the comparisons in this video are a vendor’s own comparison of itself against a rival.

Three kinds of number appear and they are not sourced the same way.

A published score is a figure on a leaderboard or a model page, cited here at the page that carries it. A vendor comparison is one lab’s claim about another lab’s model, and is cited to the lab that made the claim rather than presented as neutral. An illustration is arithmetic the video does itself out of published per token rates; it is not a measured price for any real piece of work, and the script says so where it appears.

Prices are the standard API rates for requests under the long context threshold, as listed on 23 September 2026. Several of them are introductory or promotional and will move.

The two releases, and the week they landed in

The script’s “sol launched on september twenty second, joining astra, which arrived earlier in the month” and “anthropic released opus 5.5 the same day as sol” are both correct.

Prices, per million tokens

All from OpenAI’s own pricing table, standard processing, short context. https://developers.openai.com/api/docs/pricing

model input cached input output
gpt-6-astra $10.00 $1.00 $50.00
gpt-6-sol $2.00 $0.20 $10.00
gpt-6-luna $0.10 $0.01 $0.50
gpt-5.6-sol $4.00 $0.40 $20.00

Claude opus 5.5 is $4.00 input, $20.00 output, $0.20 cache read and $5.00 cache write. https://www.anthropic.com/claude-opus-5-5

Gemini 3.8 flash is $0.75 input and $3.75 output at its introductory rate, moving to $1.50 and $7.50 on 1 January 2027. https://ai.google.dev/gemini-api/docs/latest-model

Every ratio the script states follows from that table:

Context window and the long context threshold

Prompt caching

Cached input is priced at one tenth of standard input on every GPT 6 model, which is the ninety per cent discount the script states: astra $10 to $1.00, sol $2 to $0.20, luna $0.10 to $0.01. Output is not discounted at any cache rate, and cache writes carry their own price, listed separately, at $12.50 per million for astra. https://developers.openai.com/api/docs/pricing

The worked example

This is the video’s own arithmetic on the published rates, not a measured bill, and the script says so where it appears.

At 100,000 uncached input tokens and 20,000 output tokens, standard short context rates:

Token charges only. It excludes tool calls, cache writes, and every retry a real job takes, which is the point the script makes immediately afterwards.

FrontierCode

Cognition’s benchmark for whether a change is mergeable, not merely whether it passes. Tasks are written by the open source maintainers of the repositories they come from, and submissions are scored for correctness, tests, scope, style and maintainability against maintainer authored rubrics. https://cognition.com/blog/frontier-code https://cognition.com/frontiercode

OpenAI reports that GPT 6 sol matches Claude fable 5.1 at extra high effort on FrontierCode at substantially lower cost per task. https://openai.com/index/introducing-gpt-6-sol-and-luna/

That is a vendor comparison and it is the form the script uses. Sol’s own FrontierCode percentage is not put on screen in this video: Cognition’s public leaderboard loads its table dynamically and the GPT 6 entries could not be read off it directly, and the figures circulating in secondary write ups could not be confirmed against Cognition’s own page. The claim that survives sourcing is the one OpenAI makes, so that is the one the picture carries.

DeepSWE v1.1

Datacurve’s benchmark: 113 long horizon engineering tasks across 91 repositories and five languages, with agents committing changes that are then verified in isolated containers so the environment cannot be gamed. https://deepswe.datacurve.ai/blog/deepswe-v1-1

Leaderboard, pass@1 with the stated uncertainty and average cost per task:

model pass@1 avg cost per task
GPT 6 astra 74% ±3% $4.43
Gemini 3.8 flash 74% ±1% $2.36
Claude opus 5 74% ±4% $11.84
GPT 5.6 sol 73% ±3% $6.46
Claude fable 5 70% ±3% $13.41
GLM 5.3 69% ±3% $3.99
Kimi K3 69% ±5% $4.65

https://deepswe.datacurve.ai/blog/deepswe-v1-1

Terminal Bench 4.0 and Terminal Bench Science 0.1

Anthropic publishes both tables with its competitors’ scores beside its own. https://www.anthropic.com/claude-opus-5-5

model Terminal Bench 4.0 Terminal Bench Science 0.1
Claude opus 5.5 66.4% 58.7%
GPT 6 astra 57.9% 64.6%
Claude fable 5.1 55.8% 52.6%
Claude opus 5 52.3% 29.0%
GPT 5.6 sol 37.3% 22.4%

The fallback models caveat

Anthropic states it directly: “Claude Opus 5.5 was evaluated with its production safeguards enabled. When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5.” https://www.anthropic.com/claude-opus-5-5

That is what the script means by the delivered setup rather than an isolated contest between two raw models, and it is Anthropic’s own disclosure rather than an outside criticism of it.

AutomationBench 1.0.6

Zapier’s benchmark. Each task boots a simulated company across 47 simulated business tools spanning sales, marketing, operations, support, finance and HR, hands the agent a request, and then grades the resulting state of that environment rather than the agent’s own account of what it did. There is no model acting as judge. Scoring is strict pass or fail per task, 1.0 only if every assertion passes, and the pass rate is the average across tasks. https://zapier.com/benchmarks https://zapier.com/blog/introducing-automationbench/

model pass rate
GPT 6 astra 41.4%
Claude opus 5.5 40.0%
GPT 6 sol, extra high effort 33.2%
Claude fable 5.1 31.4%
GPT 5.6 sol 28.8%
Claude opus 5 26.9%

Zapier’s own leaderboard also publishes a cost per task column: astra $1.73, opus 5.5 $1.28, and sol $0.27. https://zapier.com/benchmarks

https://www.anthropic.com/claude-opus-5-5 https://www.vellum.ai/blog/gpt-6-sol-and-luna-benchmarks-explained

This is the “published automation test” of the opening line. Sol’s 33.2% does beat Claude opus 5’s 26.9%, and sol runs the suite at $0.27 per task, which OpenAI reports as roughly nine per cent of what Claude opus 5 spends on it. https://www.vellum.ai/blog/gpt-6-sol-and-luna-benchmarks-explained

The nine per cent is OpenAI’s own comparison. Claude opus 5’s per task cost on this suite is not separately published, so the ratio rests on OpenAI’s reporting rather than on two figures that can be divided by each other here. Sol’s own $0.27 is published.

The script’s later point that sol sits below opus 5.5 and astra on the same suite is the same table read honestly: 33.2 against 40.0 and 41.4. The opening compares sol with opus 5, and the later chapter compares it with opus 5.5, which is a different model released three weeks later.

Generation speed

Artificial Analysis measures GPT 6 sol at about 126 output tokens per second and GPT 6 astra at about 53, roughly a 2.4 ratio, which is the script’s “more than twice the rate”. https://artificialanalysis.ai/models/releases/comparisons/gpt-6-sol-vs-gpt-6-astra

Two details on that page are worth stating exactly, because they are close to the script’s wording without matching it. Its own summary line reads “For output speed, GPT 6 Sol is fastest: GPT 6 Sol (max) at 125 t/s, against GPT 6 Astra (xhigh) at 53 t/s.” So the sol figure is 125 rather than 126, and the astra figure is listed at extra high effort rather than at maximum. These are live measurements that move between samples, and the script says “about” on both numbers; the ratio is what the beat is actually about. Because the page’s own wording differs from the narration, the effort labels are kept off screen on that beat and the drawing carries the two rates and the credit.

This measures generation only. It does not measure reasoning time, tool calls, or the rounds an agent spends correcting itself, which is the distinction the script draws immediately afterwards.

Changing the job while it is running

Both of those are the basis for the script’s point that a new requirement belongs inside the current job rather than forcing a restart.

Subscriptions against API billing

Sol reaches ChatGPT Work and Codex through plan tiers, and the API separately on metered per token rates. https://openai.com/index/introducing-gpt-6-sol-and-luna/ https://developers.openai.com/api/docs/pricing

These are two different billing systems, which is the script’s point that an API price cut does not automatically halve a subscription. No figure is stated for any subscription and none is put on screen.

Not checked