# Try GPT-5.6 Sol for coding this afternoon > Source: > Published: 2026-07-10 00:00:00+00:00 ## Sol lands a point behind Fable 5 at a third of the price OpenAI released GPT-5.6 Sol on 9 July 2026, alongside two cheaper siblings. On the benchmark firm’s Intelligence Index — an aggregated score across reasoning, math, coding and agentic tests — Sol scored 59, one point behind Claude Fable 5’s 60 and ahead of Claude Opus 4.8 (56), GPT-5.6 Terra (55) and the older GPT-5.5 (55), [per the Artificial Analysis run reported by The Decoder](https://the-decoder.com/gpt-5-6-sol-nearly-matches-fable-5-on-aggregated-benchmarks-at-one-third-the-cost/). The interesting number isn’t the score — it’s the cost. Per-task pricing in the same run puts Sol at about $1.04, against Fable 5’s $2.75. Artificial Analysis says Sol now sits at the best-yet trade-off between score and output tokens: almost-Fable 5 quality, one-third the bill, and Sol burns up to 54% fewer output tokens on agentic coding work than comparable models, [according to OpenAI CEO Sam Altman](https://the-decoder.com/gpt-5-6-sol-nearly-matches-fable-5-on-aggregated-benchmarks-at-one-third-the-cost/). ## The coding-agent chart is where Sol pulls ahead The story shifts on the Coding Agent Index. Run inside OpenAI’s Codex environment, Sol takes the top spot at 80 points — beating Fable 5 in Claude Code (77), Terra in Codex (77), and the older GPT-5.5 in Codex (76). On the [AA-Briefcase benchmark for realistic office tasks](/articles/aa-briefcase-a-tougher-test-for-agents/), Sol earns the highest score on slide-deck-style work, though Fable 5 still leads the overall ranking. Independent testers are seeing the same shape. Claire Vo, who runs ChatPRD and the *How I AI* podcast, put Sol through a five-category benchmark of her own design — product specs, prototypes, wireframes, debugging, and agentic voice — against Fable 5, Sonnet 5, Terra and Luna. Sol won her weighted index by a meaningful margin, [she wrote in her newsletter](https://www.lennysnewsletter.com/p/gpt-56-sol-vs-claude-fable-why-openais). Vo’s verdict: Sol breaks through where Fable 5’s precision and pedantry get stuck on real product work. The community’s reaction has tracked the cost story more than the score. The independent benchmarking account @zefirium noted Sol’s 91.9% on Terminal-Bench 2.1 and its Ultra mode, where the model spawns cooperative subagents that share context mid-task: GPT-5.6 is here. Three tiers, one government clearance, and a pricing war that just reshaped the entire AI market: Sol ($5/$30): flagship. Ultra mode spawns cooperative subagents sharing context mid-task. 91.9% Terminal-Bench 2.1. Terra ($2.50/$15): GPT-5.5 performance at… — ＺＥＦＩ (@zefirium)[Jul 9, 2026] ## Try it this afternoon For a UK team running a coding agent on real work — code review, ticket triage, weekly reporting — the play this week is a controlled switch, not a leap. The cheapest credible route to near-Fable 5 work, run today, is GPT-5.6 Sol inside Codex. A simple, costed experiment: **Pick one workflow you’ve already costed.** A weekly report, a refactor task, a triage queue. You want a baseline you trust.**Run it on Fable 5 (or whatever you’re on) and on Sol.** Keep the prompt identical, log the token spend on each side.**Lean on the cache-read discount.** Sol charges a small fee the first time a piece of context is loaded, then drops the price by 90% for any later re-use — a clear nudge towards long agent sessions that re-read the same files.**Cap the loop.** A 10-minute wall-clock and a token budget prevent the cost maths from drifting mid-test.**Score the output yourself.** Benchmark numbers are an average; your work is the one that pays the bills. For a team running 1,000 agent tasks a month, the per-task delta alone — $1.04 vs $2.75 — saves around £1,350 a month at current exchange rates. Luna, the cheapest tier, lands at $0.21 a task, but expect Fable 5-class quality only on simpler briefs. If you’d rather not re-platform, proxies like Sage Router let you A/B models through a single endpoint — see [our routing write-up](/articles/sage-router-one-endpoint-every-model/). Three caveats before you switch: **Restricted US access is still in force.**[Our earlier coverage of the Sol launch](/articles/openai-ships-gpt-5-6-sol-under-restricted/)flagged government-cleared tiers and regional limits. UK access is available via the standard OpenAI API and Codex, but check your account tier before assuming Ultra mode is on.**The cost story depends on the harness.** Sol’s per-task figure assumes Codex. Run it through a third-party wrapper and the token economics shift.**Luna is not a free Fable 5.** At $0.21 a task it’s tempting, but expect noticeable quality drops on long, multi-step briefs. Use it for the cheap, well-scoped jobs. If you’ve been holding off on a coding agent while waiting for the price-vs-quality curve to settle, the curve just settled a notch. ## Sources & quotes Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. [How we verify →](/blog/how-we-keep-an-ai-newsroom-honest/)