GPT-5.6 Sol Tops the Coding Leaderboard: At What Cost?

Number one has an asterisk. As of July 9, 2026, OpenAI’s GPT-5.6 reached general availability, the government access gate came off, and its flagship tier, Sol, took the top spot on the Artificial Analysis coding-agents leaderboard. On the independent Coding Agent Index it scores 80, about 2.8 points clear of Claude Fable 5 at 77.2. First on the board that matters most for agentic coding.

Then you read the next column. Fable 5 still wins SWE-Bench Pro by a wide margin, 80.3 percent to Sol’s 64.6, and edges it on general intelligence. The eye-catching 91.9 percent everyone is quoting is Sol in Ultra mode, which runs four agents in parallel and bills you for all four. So the lead is real. It is also narrow, split, and at the very top, expensive.

Here is the part a leaderboard screenshot leaves out: which number you cite is the argument you are making. This is what actually changed at GA, where Sol wins and where it does not, and what the headline score costs per finished task.

The board after general availability

The tier structure is not the news. OpenAI previewed Sol, Terra and Luna back on June 26, with the prices that still hold today. What was missing then was two things: access, and independent numbers. Both arrived at GA.

Access first. The June preview shipped into a limited rollout of roughly 20 government-approved partners, the arrangement a June executive order set up as a release gate. On July 9 that gate lifted. GPT-5.6 is now live across ChatGPT, Codex and the API, with the three tiers billed at list price, as reported by MarkTechPost. GitHub turned the same three tiers on inside Copilot the same day.

The second arrival matters more for anyone deciding what to run: independent numbers. Until GA, every GPT-5.6 score was OpenAI’s own. Now Artificial Analysis has run the models through its own harness, and its ranking is the one the “Sol is number one” claim actually rests on.

That ranking is the Artificial Analysis Coding Agent Index, a composite of three agentic evals: DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA. Sol tops it.

Model	AA Coding Agent Index v1.1
GPT-5.6 Sol (max)	80
GPT-5.6 Terra	77.4
Claude Fable 5	77.2
GPT-5.6 Luna	74.6

Note that Terra, the mid tier, edges Fable 5 too. Artificial Analysis has Sol leading all three of the index’s sub-evals, tying Grok 4.5 on SWE-Atlas-QnA, and it reaches the top score using less than half the output tokens and less than half the time of Fable 5, at about a third less cost. The 2.8-point margin over Fable 5 is the headline. The efficiency behind it is the part that changes a bill.

The single most-quoted line item inside that index is Terminal-Bench v2.1, and it is worth pulling out on its own, because the same benchmark shows different scores depending on who ran the eval.

Model and mode	Terminal-Bench v2.1	Score source	Price /1M (in / out)
Sol, Ultra (4 agents)	91.9%	OpenAI, self-reported	$5 / $30
Sol, xhigh reasoning	89.5%	Artificial Analysis, independent	$5 / $30
Sol, single (max)	88.8%	OpenAI, self-reported	$5 / $30
Sol, max	88.0%	Artificial Analysis, independent	$5 / $30
Terra, max	88.0%	Artificial Analysis, independent	$2.50 / $15
Luna, single	84.7%	OpenAI, self-reported	$1 / $6
Claude Fable 5, Claude Code	83.1%	Terminal-Bench board, independent	$10 / $50

Terminal-Bench itself is independent: an open benchmark maintained by Stanford and the Laude Institute, run in isolated sandboxes against real test suites. The “score source” column does not question that. It tracks who executed the run and published the number. A vendor can run the public benchmark on its own model and report the result (self-reported, the OpenAI rows here), or a neutral party can run it: Artificial Analysis in its own harness, or the official Terminal-Bench leaderboard. Self-reported is not the same as wrong. But it uses the vendor’s chosen harness, reasoning effort, and retry policy, so the independent rows carry more weight. The reassuring sign here is that OpenAI’s self-reported 88.8 for a single agent lands right next to Artificial Analysis’s independent 88.0 to 89.5.

Sol vs Fable 5: two leaderboards, two winners

Put the two flagships side by side across the benchmarks that decide coding work and the picture is not a ladder. It is a two-two split. Sol pushes out on the agentic axes, the ones that reward driving a terminal and finishing a long task. Fable 5 pushes out on hard codebase resolution and general reasoning.

GPT-5.6 Sol vs Claude Fable 5 capability fingerprint

Axis	Claude Fable 5	GPT-5.6 Sol
Coding Index	77.2	80.0
SWE-Bench Pro	80.3	64.6
Terminal-Bench	83.1	88.0
Intelligence	59.9	58.9

SWE-Bench Pro is the axis that should give a buyer pause. It grades whether a model can resolve real, hard issues inside a large codebase, and Fable 5 holds a wide lead there, 80.3 percent on Anthropic’s own launch benchmarks against Sol’s OpenAI-reported 64.6. Both are vendor self-reported on the model maker’s own scaffolding, and independent aggregators contest Anthropic’s number, so read the gap as a claim, not a neutral measurement. Terminal-Bench rewards something different: can the agent drive a shell, chain commands, and not get lost across a long session. Sol owns that. If your work looks like agentic operations and long tool-use chains, Sol’s shape is the one you want. If it looks like landing correct patches in a big, messy repository, Fable 5 still has the edge, which is the cost side of the same question covered in GPT-5.6 vs Claude Fable 5 cost per task. What each of these tests actually measures is broken down in the AI benchmarks directory.

The Intelligence Index gap is small, 59.9 to 58.9, and points the same way: Fable 5 is marginally the stronger generalist, Sol the stronger agent. One point on a composite is inside the noise. The SWE-Bench Pro gap is not.

The 91.9 percent only happens in Ultra mode

Ultra is a mode inside GPT-5.6 Sol, not a separate model. It splits a single task across four agents running in parallel by default, each working one piece before the results are synthesized. OpenAI reports Ultra at 91.9 percent on Terminal-Bench v2.1 against 88.8 for a single Sol agent: a gain of 3.1 points for four times the agents.

Every quote of GPT-5.6 as a 91.9-percent model is quoting Ultra, and it is easy to misread as a fourth model or a free upgrade. OpenAI says the subagents are trained to coordinate mid-task rather than merge only at the end, which is what separates Ultra from running the same prompt four times.

Set the numbers next to each other and the shine comes off. A single Sol agent already reaches 89.5 percent in independent testing at xhigh reasoning. Ultra’s self-reported 91.9 sits about 2.4 points above that, bought with roughly four times the token spend. At the frontier the returns bend hard: you pay a multiple for the last couple of points, which is exactly where the site’s cost-per-task math tends to punish the fanciest configuration.

What the top score costs per task

The pricing is the quiet win. Sol lists at $5 input and $30 output per million tokens, exactly GPT-5.5’s old rate, so OpenAI raised capability without raising the flagship’s sticker. Terra runs $2.50 and $15, Luna $1 and $6. Against Fable 5 at $10 and $50, Sol is roughly half the per-token cost.

Per finished task, the gap is wider than the token rate suggests. Artificial Analysis puts Sol at about $1.04 per task on its Intelligence Index, using around 15,000 output tokens per task, less than half the output of most models at its level, at roughly a third of Fable 5’s cost. That is the number that actually moves a monthly bill: not the rate card, the tokens burned to reach an answer.

Ultra inverts that advantage. Four agents on one task means roughly four times the tokens, so the mode that wins the headline is the mode that erases the cost edge. For the 3.1 points, most teams should not pay it by default. Reserve Ultra for the small set of tasks where a wrong answer is expensive enough to justify the multiple, and route everything else to a single Sol or a Terra agent. Drop your own token counts into the cost-per-task calculator to see where the line sits for your workload.

Where Sol sits in the wider field

One board does not settle a field. Sol leads the coding-agents index today, but the full value leaderboard ranks models by score per dollar, and cheap open-weight models keep winning that fight on routine work. The 2026 AI coding-agent landscape puts Sol in context against Claude Code, Cursor, Copilot and the rest, where the agent harness often matters more than the model inside it. That harness effect is the whole finding of the OpenCode vs Claude Code vs OpenClaude benchmark run: same model, different wrapper, different score.

Two GPT-5.6 stories sit alongside this one and are worth the click if they touch your stack. The government release gate explains why access was rationed for two weeks and why that favors incumbents. The Cerebras speed story covers Sol at up to 750 tokens per second, which changes what the model is usable for without changing what it costs per token.

Bottom line

Sol genuinely tops the Artificial Analysis coding-agents leaderboard after GA, and it does it while holding GPT-5.5 pricing and burning fewer tokens per task than anything at its level. That is a strong release. But the number-one claim is a 2.8-point margin on one index, Fable 5 still wins hard codebase resolution and edges general intelligence, and the 91.9 percent that made the headlines is a four-agent Ultra self-report you pay four times over. Rank by the independent single-agent scores, decide by cost per finished task, and treat Ultra as a deliberate spend on the tasks that earn it. First place, read honestly, is a recommendation to run Sol for most agentic work, not a reason to always run it flat out.

Frequently asked questions

Does GPT-5.6 Sol top the coding-agents leaderboard?

Yes. After reaching general availability on July 9, 2026, Sol leads the independent Artificial Analysis Coding Agent Index at 80, about 2.8 points above Claude Fable 5 at 77.2. It also leads Terminal-Bench v2.1. The lead is narrower on other tests.

What is GPT-5.6 Ultra mode?

Ultra is a mode inside the Sol model, not a separate GPT-5.6 model. It decomposes a task and runs four agents in parallel by default, then synthesizes their work. OpenAI reports Ultra at 91.9 percent on Terminal-Bench v2.1 versus 88.8 percent for a single Sol agent, at roughly four times the token cost.

Is GPT-5.6 Sol better than Claude Fable 5?

It depends on the task. Sol leads the Artificial Analysis Coding Agent Index and Terminal-Bench, so it is stronger on agentic and terminal work. Fable 5 leads the vendor-reported SWE-Bench Pro 80.3 percent to 64.6 and edges the independent Intelligence Index 59.9 to 58.9, so it is stronger on hard codebase resolution and general reasoning. Sol also costs about half as much per token.

What is the Terminal-Bench v2 leaderboard?

It ranks coding agents by their pass rate on Terminal-Bench v2.1, a set of hard command-line tasks run in isolated sandboxes. Independent Artificial Analysis results put Sol near 88 to 89.5 percent and Terra near 88 percent, which differ from OpenAI self-reported figures because the harness and settings differ.

How much does GPT-5.6 Sol cost?

Sol is $5 per million input tokens and $30 per million output, the same as GPT-5.5. Terra is $2.50 and $15; Luna is $1 and $6. Artificial Analysis measured Sol at about $1.04 per task on its Intelligence Index, roughly a third of Fable 5. Ultra mode multiplies the bill by about four because it runs four agents.

Is GPT-5.6 Sol Ultra worth the extra cost?

For most work, no. Ultra buys about 3.1 self-reported benchmark points over a single Sol agent for roughly four times the token spend, and a single agent already scores near 89.5 percent independently. Reserve Ultra for tasks where a wrong answer is expensive enough to justify the multiple.

Sources

• OpenAI. (2026). GPT-5.6: Frontier intelligence that scales with your ambition. Primary announcement; the July 9 2026 general-availability release, tiers, Ultra mode, pricing, and self-reported benchmarks. Verified 2026-07-10. openai.com/index/gpt-5-6
• Anthropic. (2026). Claude Fable 5 and Claude Mythos 5. Vendor announcement; Fable 5 launch benchmarks including a self-reported SWE-Bench Pro 80.3 percent, run on Anthropic scaffolding and contested by independent aggregators. Verified 2026-07-10. anthropic.com/news/claude-fable-5-mythos-5
• Artificial Analysis. (2026). GPT-5.6 has landed. Independent evaluation; Coding Agent Index (Sol 80, Terra 77.4, Luna 74.6, Fable 5 77.2), Intelligence Index, and per-task cost. Verified 2026-07-10. artificialanalysis.ai/articles/gpt-5-6-has-landed
• Artificial Analysis. (2026). Terminal-Bench v2.1 evaluation. Independent leaderboard; Sol 88.0 percent (max) and 89.5 (xhigh), Terra 88.0. Verified 2026-07-10. artificialanalysis.ai/evaluations/terminalbench-v2-1
• MarkTechPost. (2026). OpenAI releases GPT-5.6 (Sol, Terra, Luna): a three-tier model family with programmatic tool calling in the Responses API. Secondary coverage; GA access details, OpenAI self-reported benchmark table, Ultra definition (four parallel agents), pricing. Verified 2026-07-10. marktechpost.com
• GitHub. (2026). OpenAI’s GPT-5.6 Sol, Terra and Luna are now available in GitHub Copilot. Primary changelog; the three tiers, availability by SKU, list-price billing, July 9 2026. Verified 2026-07-10. github.blog/changelog
• Terminal-Bench. (2026). Terminal-Bench 2.1 leaderboard. Independent benchmark maintained by Stanford University and the Laude Institute; Claude Fable 5 in Claude Code at 83.1 percent. Verified 2026-07-10. tbench.ai/leaderboard/terminal-bench/2.1