OpenAI shipped GPT-5.6 Sol on July 9th, and if you blinked, you missed the most important line in the entire release. It is not the intelligence score. It is not the agentic coding result. It is the fact that Anthropic is now selling the most expensive model in its class at three times the price of an almost-identical competitor, and Sam Altman knows it.
I have been staring at the Artificial Analysis Intelligence Index since the numbers dropped, and the headline everyone is quoting is wrong. Fable 5 leads with 60. Sol has 59. That gap is a rounding error. What is not a rounding error is the cost column, and that is the part of the table nobody wants to print in full.
Here is the entire pricing floor, pulled from Artificial Analysis and OpenAI's release notes.
Intelligence Index vs Cost per Task
| Model | Intelligence Index | Cost per Task | Provider |
|---|---|---|---|
| Claude Fable 5 (max) | 60 | $2.75 | Anthropic |
| GPT-5.6 Sol (max) | 59 | $1.04 | OpenAI |
| Claude Opus 4.8 | 56 | (higher) | Anthropic |
| GPT-5.6 Terra | 55 | $0.55 | OpenAI |
| GPT-5.5 | 55 | (mid) | OpenAI |
| GPT-5.6 Luna | (lower) | $0.21 | OpenAI |
Sol sits one point behind the leader and costs roughly a third of the price. Terra sits four points behind Sol and costs half as much again. Luna is 80% cheaper than Sol, and it is still a GPT-5.6. The Pareto frontier Artificial Analysis bragged about is real, and it is tilted hard against Anthropic.
Coding Agent Index: Sol Actually Wins
If the Intelligence Index is a near-tie, the agentic coding result is not even close.
| Model | Coding Agent Score | Environment |
|---|---|---|
| GPT-5.6 Sol | 80 | Codex |
| GPT-5.6 Terra | 77 | Codex |
| Claude Fable 5 | 77 | Claude Code |
| GPT-5.5 | 76 | Codex |
Sol does not "nearly match" Fable 5 here. It tops the chart, three points clear. If you are a developer running real agent loops and you are still defaulting to Fable 5 because it is the safe name to drop in a pull request description, you are paying a 3x premium for three fewer points on a benchmark and a brand name.
The Cache-Write Fee Is the Real Story
The per-million-token headline numbers look almost reasonable. Sol is $5 input and $30 output. Fable 5 is $10 and $50. You double it, you win. But OpenAI added something Anthropic has not: a cache-write fee.
The first time you write a long prompt into Sol's prompt cache, you pay a one-time cache-write cost. Every subsequent read is 90% off. If your workflow hits the cache, your effective price collapses further. If it does not, you are still cheaper than Fable 5 on raw input.
This is the part of the release OpenAI buried on page three and Anthropic will never quote. It is also the part that matters for any team running repeated agent loops against the same codebase.
What Anthropic's "Advisor" Pattern Admits
I want to be specific about this, because the timing is damning. Anthropic published their own cost mitigation the same week: run Fable 5 only as a planner, delegate execution to Sonnet 5.
# Anthropic's published "Advisor" pattern
planner = claude_fable_5 # expensive, only for planning
executor = claude_sonnet_5 # cheap, does the work
plan = planner.think(task)
result = executor.run(plan)That setup hits 92% of Fable 5's solo performance at 63% of the cost, according to Anthropic's own numbers. Read that again. Anthropic, the company that built Fable 5, is publicly telling you to use it less. The recommended pattern is "use the expensive model to tell the cheap model what to do." OpenAI shipped a model that just is the cheap model, and gets within one point.
That is not a benchmark gap. That is a product positioning crisis.
The 54% Token Efficiency Claim
Sam Altman told CNBC that Sol is "54% more token efficient on agentic coding" than its predecessor. That number is doing a lot of work, but the direction is not in dispute. The Artificial Analysis page itself credits Sol with "fewer output tokens than models with similar performance, up to 54% fewer for agentic coding tasks." When you combine lower cost per token and fewer tokens per task, the math stops being close.
What This Means If You Are Picking a Default
I am not going to tell you to switch off Fable 5 today. There are workloads where Anthropic's tool use is still cleaner, and there are teams whose entire infra is pinned to Claude Code. But the next time someone in your org asks "why are we paying $10/M input tokens," the honest answer is no longer "because it is the best." It is now "because it is one point better, and we have not re-priced our defaults since Sol shipped."
Three things I would actually do this week:
- Re-run your top three agentic coding workloads on Sol and measure cost per completed task, not per token. The token number lies.
- If you are on Anthropic, try the Advisor pattern. Even if you keep Fable 5 in the loop, you will cut your bill by a third.
- Pressure your vendor for a price match. OpenAI just moved the floor. Quietly paying $10/M is now a strategy choice, not a default.
The Take Nobody Will Print
This is the week Anthropic stopped being the obvious safe pick. The Intelligence Index gap is gone, the Coding Agent Index is lost, and OpenAI is selling the new winner for a third of the price with a cache discount on top. Anthropic will respond. They have to. But until they do, every team that buys on price-per-quality is already migrating, and every team that does not is paying a tax to feel safe.
Source: Artificial Analysis Intelligence Index and Coding Agent Index, July 9 2026.