GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price
OpenAI’s release of GPT‑6.1 Sol, a “near‑Astra” model at roughly one-fifth Astra’s price, is being met with both optimism over cheaper high-end capability and skepticism about real-world gains over earlier models and Anthropic’s Opus 5.5. Commenters focus heavily on economics: aggressive API price cuts, much cheaper cached tokens, and simultaneous reductions in subscription allowances are seen as signs of intensifying price competition, looming IPOs, and pressure from strong Chinese and open‑weight models. Many report that recent OpenAI releases (especially GPT‑6 Sol) felt like regressions for coding and agentic work, reinforcing a sense that core intelligence is improving more slowly while vendors now compete mainly on cost, efficiency, and harness quality.
Perceived quality of GPT‑6.1 Sol vs 6.0, 5.6, Astra, and Opus 5.5
- Many say GPT‑6.0 Sol was a clear regression from 5.6 Sol: more dumb mistakes, over‑engineering, poor code, needed “babysitting”; several reverted to 5.6 or switched to Opus.
- 6.1 Sol is widely seen as a “fix” or replacement: described as near‑Astra in capability at much lower cost, and likely derived from Astra or an “Astra Light” variant.
- Benchmarks and some user tests show 6.1 Sol close to Astra and competitive with Opus 5.5, especially on coding/math tasks, at lower per‑task cost.
- However, a large contingent still finds Opus 5.5 clearly better in real projects (complex coding, reasoning, “taste” in code/prose), and uses Astra mainly for vision/3D/Blender tasks.
Pricing, subscriptions, and business model
- API token prices and especially cache pricing are praised (notably $0.10/M cached tokens and overall DeepSeek/Anthropic parity).
- At the same time, the $200 Pro plan is being cut from 20× to 10× usage, while a new $500 plan appears; this is widely viewed as a “nerf” and push toward higher tiers or pure API.
- Several users say Claude subscriptions currently deliver more usable compute per dollar; others note that frontier models are expensive to serve and OpenAI is likely losing money on subs.
Release cadence, “slowing down,” and safety talk
- GPT‑6.1 Sol arrives only a week after 6.0 Sol; many see this as a rushed response to negative feedback on 6.0 and to Opus 5.5, not “pacing the frontier.”
- Some argue all the recent pricing and coordination rhetoric is about compute and token‑demand limits, not safety; others think capability growth is still clearly ongoing, not plateaued.
Benchmarks vs real‑world usage
- Several benchmark suites show 6.1 Sol matching Astra and roughly tying or slightly beating Opus 5.5 at similar or lower cost.
- Many engineers distrust these numbers: they report that models can benchmark well but still underperform on large legacy codebases, refactors, and long agentic workflows.
- There’s repeated frustration that models make code more verbose, add needless abstractions, and vary in instruction‑following despite strong scores.
Competition from open‑weight and Chinese models
- DeepSeek, Qwen, GLM, Kimi, and MiMo are heavily discussed. Users like their extremely low prices and say they’re “good enough” for most CRUD and routine coding.
- Others note they often consume far more tokens (especially reasoning tokens), hallucinate more, or feel “brilliant but stupid,” so frontier models can still be cheaper and better per completed task.
Harnesses, agents, and token burn
- Experience varies dramatically depending on harness (Codex, Claude Code, OpenCode, Pi, OMP) and subagent strategies.
- Heavy use of subagents and frequent compaction can burn through weekly quotas in hours; careful orchestration, context limits, and use of cheaper sub‑models are seen as crucial.