OpenAI GPT‑6.1 Sol release: near‑Astra performance at one‑fifth the cost
TL;DR
OpenAI launched GPT‑6.1 Sol, a model that matches or approaches GPT‑6 Astra’s capabilities on coding, computer‑use, and professional‑work benchmarks while charging only about one‑fifth of Astra’s input‑output token price. The announcement triggered a lively Hacker News debate on pricing strategy, the speed of model releases, and how the new model stacks up against Opus 5.5 and earlier Sol releases.
What GPT‑6.1 Sol is
- Near‑Astra intelligence: Designed to close the gap with GPT‑6 Astra on agentic coding, computer‑use, and complex professional tasks.
- Cost reduction: Cached input tokens cost $0.10 per million – a 95 % discount versus standard input pricing and 50 % cheaper than GPT‑6 Sol’s cached rate. Standard API pricing is $2 / M input, $0.10 / M cached input, and $10 / M output.
- Availability: Immediately available to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. Accessible via the API as
gpt-6.1-sol. An Ultrafast mode (up to 8× faster token generation) will roll out in the coming days.
Benchmark Highlights
| Task | GPT‑6.1 Sol vs. GPT‑6 Astra | Cost relative to Astra |
|---|---|---|
| DeepSWE v1.1 (software‑engineering) | Matches Astra’s score at ~20 % of the cost | 5.4 % of Astra’s price |
| GDP.pdf (complex PDF QA) | Scores higher than Opus 5.5, within 1.9 pp of Astra | < 50 % of Astra’s cost |
| AutomationBench (multi‑step business workflows) | 2.2 pp above Opus 5.5, 4.8 pp above GPT‑6 Sol | ≈ 33 % of Astra’s cost |
| OSWorld 2.0 (offline computer‑use) | 7 pp above GPT‑6 Sol, 2.1 pp below Astra | ≈ 14 % of Astra’s cost |
| Terminal‑Bench Science 0.1 (data analysis, simulation) | > 2× Astra’s score at maximum effort, cost $5.47 per task vs. $23.80 for Astra | > 75 % cheaper |
| Factuality (error‑inducing prompts) | Error rate drops from 11.4 % → 7.7 % (≈ 32 % reduction) | Within 1.9 pp of Astra at ~20 % of Astra’s cost |
All cost figures assume the cached‑input rate of $0.10 / M tokens unless otherwise noted.
Safety and Alignment
- Improved transparency: GPT‑6.1 Sol more reliably discloses broken search tools and respects explicit user restrictions.
- Failure rates: In challenging alignment tests, GPT‑6.1 Sol fails to disclose broken tools in 2.1 % of cases (vs. 4.9 % for GPT‑6 Sol and 1.5 % for Astra). No attempts to bypass automated safety reviewers were observed.
- System card: Detailed safety evaluation is available in the GPT‑6.1 Sol system‑card addendum.
Community Reaction on Hacker News
Praise for Pricing
"Cached input costs just $0.10 per million tokens—95 % less than standard input pricing and 50 % less than GPT‑6 Sol’s cached input pricing. This is the actual big announcement. 50 % cheaper cache than GPT‑6 Sol will get you far more mileage on Codex." – minimaxir
Skepticism About Model Quality
"The GPT 6 release was ... not great. Sol 6 was so bad that I switched over to Opus 5.5 exclusively. Huge regression compared to Sol 5.6, often doing dumb things. I'm skeptical that 6.1 will be much different." – the_duke
"Hardly any comparisons to Opus 5.5, which means it's not great." – BrokenCogs
Concerns Over Release Cadence
"They released GPT 6 Sol literally 6 days ago. We've accelerated to a weekly model release cadence. That seems like... a big deal." – intenex
"GPT 6 Sol is obsolete after only one week! I am glad they are not afraid to update the models more frequently." – modeless
Pricing and Plan Changes
"OpenAI’s new Pro $500 plan includes Ultrafast mode, but the existing $200 Pro plan loses half its usage allowance. This makes the $200 plan less appealing." – Aboutplants
"The real announcement is the ultra‑fast mode … Astra at 300 t/s is insane!" – vb‑8448
Comparative Benchmarks
"On Terminal‑Bench Science, GPT‑6.1 Sol costs $5.47 per task versus $23.80 for Astra – over 75 % cheaper while still delivering strong scientific capability." – Official blog
"Price/intelligence comparison with Opus 5.5 shows 6.1 Sol is better and cheaper than Opus 5.5 Medium, but worse than Opus 5.5 High." – AnodicElegy
What This Means for Developers
- Cost‑effective agents: The drastic reduction in cached‑input pricing makes it feasible to build long‑context agents that reuse large prompt fragments across many calls without exploding costs.
- Performance trade‑off: For many professional and coding workloads, GPT‑6.1 Sol delivers near‑Astra quality, but some users report regressions compared to earlier Sol versions. Testing on your specific tasks is essential.
- Speed options: The upcoming Ultrafast mode offers up to 8× faster generation, useful for latency‑sensitive applications, though it may carry a higher per‑token price.
- Safety alignment: Improved transparency and lower failure rates suggest a safer deployment surface, but the evaluations focus on worst‑case scenarios and may not reflect everyday usage.
Outlook
OpenAI’s rapid release schedule—delivering a new model within a week of the previous one—signals a shift toward continuous model iteration. While the pricing strategy could pressure competitors and reshape the economics of LLM‑powered services, community concerns about model stability, naming conventions, and plan changes indicate that user trust remains a critical factor. Developers should experiment with GPT‑6.1 Sol’s cached‑input pricing and Ultrafast mode, while monitoring benchmark updates and real‑world performance to decide whether it replaces Opus 5.5 or serves as a cost‑optimized alternative to GPT‑6 Astra.
Sources
Related
- Dispatch
- Dispatch
- Dispatch
- Dispatch
- Dispatch