GPT-6 Sol is not the new top OpenAI model. It is the model that makes the GPT-6 generation easier to use repeatedly. Peter's launch-day comparison puts that claim into practical terms: the same max-reasoning prompts, substantially shorter waits, and in several cases far fewer total tokens than GPT-5.6 Sol.
Watch the Full GPT-6 Sol Comparison
Source and credit: GPT-6 Sol | First impressions, published by Arena AI on 22 September 2026. Peter published the test prompts in his 3D prompt collection. Reported times, token totals, and visual judgments below come from the video; OpenAI's launch page is the source of record for pricing and availability.
How Peter Tested Sol
Peter placed GPT-5.6 Sol and GPT-6 Sol side by side, gave both models the same prompt, and selected maximum reasoning for each. He then compared the generated experience, total token count, and elapsed generation time. The total includes cached tokens and other context shown by the agent interface, not merely visible output tokens.
The prompts are creative, interactive 3D worlds: balloons, Stonehenge, an octopus, environments inspired by Monet and Van Gogh, a macro raindrop scene, and a humpback whale. These tasks combine interpretation, code generation, asset decisions, rendering, testing, and browser execution. They are useful stress tests, but they are not representative of every coding or professional workload.
The Reported Runs
| Test | GPT-5.6 Sol | GPT-6 Sol | Peter's reading |
|---|---|---|---|
| Opening comparison | About 60 min | About 11 min and roughly half the total tokens | Sol produced the better result much faster |
| 3D balloons | 53 min / about 30M tokens | Under 10 min / about 4M tokens | Better generation with a major token reduction |
| Octopus | 48 min | 10 min | Sol was clearly better; token difference was smaller than the time difference |
| Van Gogh world | 54 min | About 30 min | Sol's run exposed a token-count discrepancy rather than a clean saving |
| Raindrop world | About 50 min | About 25 min | Sol was faster and visually stronger despite using more tokens in this example |
| Humpback whale | About three times longer at a similar token total | 14 min / about 6M tokens | Large latency gain without a comparable token reduction |
The balloon result is the cleanest efficiency story: roughly one seventh of the tokens and less than one fifth of the time. The raindrop and whale examples are just as important because they prevent the wrong conclusion. Faster completion does not always mean fewer tokens.
Why Wall-Clock Time and Token Use Can Diverge
Peter's theory is that older agent runs can spend a long time validating, testing, and circling a result without generating a proportionate number of tokens. Tool execution, browser rendering, tests, retries, and waiting for external processes all consume elapsed time differently from model output.
The video does not isolate those causes, so the validation theory remains an inference. The correct operational lesson is to measure both:
- Total elapsed time: from task submission to the final accepted artifact.
- Total billed usage: fresh input, cached input, reasoning, output, and tool charges.
- Agent activity: tool calls, test cycles, retries, file changes, and failed actions.
- Review burden: the minutes needed to inspect, correct, and safely accept the result.
A run can be token-efficient but operationally slow. It can also spend more tokens in parallel and still return sooner. Neither is automatically better until the output passes the same acceptance test.
Efficiency Did Not Require a Quality Collapse
Peter preferred GPT-6 Sol on the highlighted creative builds. The Monet world interpreted the artistic idea rather than placing a few generic panels. The raindrop environment created a more convincing macro perspective. The humpback whale scene had more movement and a stronger water effect. He also described the newer model as smoother and less brittle than 5.6 Sol.
He did not claim that Sol displaced the frontier tier. Stonehenge was acceptable rather than exceptional, and he judged GPT-6 Astra and Claude Fable stronger on the best visual-generation work. OpenAI makes the same hierarchy explicit: Astra remains its best model across the board.
That makes Sol's role clearer. It is the model to test when the expensive model is unnecessary but the fast model is not capable enough. The advantage is room to iterate, not a guarantee that every first pass will look frontier-grade.
Official Pricing and Caching Strengthen the Efficiency Case
| Model | Input / 1M | Output / 1M | Change |
|---|---|---|---|
| GPT-5.6 Sol | $4 | $20 | Launch comparison baseline |
| GPT-6 Sol | $2 | $10 | 50% lower list price |
OpenAI also says cached GPT-6 input receives a 90% discount and that prompt-caching changes improve default hit rates for agents and long conversations. Reasoning effort and tool availability can now change without invalidating the earlier cached context, which matters when a long coding session escalates only the difficult steps.
List price and observed total-token reductions are separate benefits. Do not multiply Peter's best token saving by the 50% price reduction and assume every task becomes fourteen times cheaper. Cache hits, output mix, reasoning settings, retries, and tool use differ across runs.
Stop Defaulting Every Task to Max Reasoning
Both sides of Peter's comparison used max reasoning, which was useful for consistency but not his recommended everyday setting. His practical advice is to try medium or high first, then escalate only for tasks that need it.
| Effort | Use it for | Escalate when |
|---|---|---|
| Medium | Routine features, visual prototypes, ordinary debugging, and familiar repositories | The model misses acceptance criteria, misunderstands architecture, or produces fragile work |
| High | Cross-file changes, ambiguous requirements, difficult bugs, and important research | A repeated high-effort run still fails for a reason deeper reasoning could plausibly solve |
| Max | Rare, expensive problems where marginal quality is worth the wait and usage | Do not escalate further; change the prompt, context, tools, decomposition, or model |
A stronger model at medium effort can be cheaper than a weaker model reasoning for an hour. The only reliable way to know is to compare accepted results, not benchmark settings.
A Five-Run Sol Evaluation
- Choose one real task. Use a task your team repeats and can score objectively.
- Freeze the environment. Keep the prompt, files, tools, permissions, and definition of done unchanged.
- Run 5.6 Sol and 6 Sol three times each. One run hides variance and luck.
- Repeat Sol at medium and high. Compare those runs with max rather than assuming the highest setting is the default.
- Record the whole task. Capture elapsed time, fresh and cached tokens, retries, tool failures, reviewer minutes, and accepted-result rate.
Adopt the lowest effort that repeatedly clears the bar. Keep Astra or another frontier model as an escalation path for the small set of tasks where Sol's lower cost creates more rework than it saves.
Video Chapters
| Time | Chapter | Time | Chapter |
|---|---|---|---|
| 00:00 | GPT-6 Sol launch and model names | 00:36 | First head-to-head comparison |
| 02:29 | 3D balloon test | 03:26 | Stonehenge comparison |
| 04:01 | Octopus test and validation theory | 05:18 | Why agent cost matters |
| 08:02 | Monet world | 09:32 | Van Gogh and token discrepancy |
| 10:01 | The efficiency trend | 11:59 | Raindrop quality uplift |
| 12:54 | Sol versus brittle 5.6 behavior | 14:21 | Try lower reasoning settings |
| 16:53 | Humpback whale and conclusion |
Sources, Prompts, and Useful Links
- Arena AI: GPT-6 Sol | First impressions
- GitHub: Peter's 3D prompt collection
- Arena AI: agent-mode comparisons
- OpenAI: Introducing GPT-6 Sol and Luna
The video and OpenAI launch material were published on 22 September 2026. This article was reviewed on 23 September 2026. Model behavior, prices, caching, availability, limits, and benchmark results can change.