Anthropic and OpenAI are locked in an increasingly tight race for frontier AI supremacy, and the gap between their flagship models are narrowing to a benchmark-by-benchmark contest.
Claude has led on real-world software-engineering tasks with Fable 5 scoring 80.3% on SWE-Bench Pro at its June launch, well ahead of GPT-5.5’s 58.6%. But OpenAI’s newly released GPT-5.6 Sol claims the state-of-the-art spot on the Artificial Analysis Coding Agent Index at 80, narrowly ahead of Fable 5, while using less than half the tokens and time.
Since launching Mythos and giving a select group of companies in its Glasswing Alliance early access, Anthropic has positioned itself at the front of the frontier race. It followed up by releasing Fable 5, a version of Mythos with extra safeguards around biology, cybersecurity, and AI research, to a much broader set of users, cementing that lead.
That momentum hit a snag when a dispute with the U.S. government over export controls led Anthropic to be treated as a supply-chain risk, forcing it to suspend access to Fable 5 and Mythos 5 for 19 days — from June 12 to July 1, 2026 — before the Commerce Department lifted the relevant controls and access was restored.
Just over a week after Fable 5 came back online for all users, OpenAI unveiled its own next-generation family: GPT-5.6 Sol, Terra, and Luna. Rather than a single flagship, OpenAI split the release into three capability tiers priced separately with Sol at the top and Luna at the bottom, arguing this lets each tier advance on its own tempo.
OpenAI has been vocal that Sol edges out Anthropic’s flagship on several fronts. On Agents’ Last Exam, an evaluation of long-running professional workflows across 55 fields, Sol scored 53.6, which OpenAI says is 13.1 points ahead of Fable 5. On the Artificial Analysis Coding Agent Index, Sol posted a new high of 80, about 2.8 points above Fable 5. It also claims new state-of-the-art marks on BrowseComp (92.2%) and OSWorld 2.0 (62.6%).
Picture isn’t one-directional
But the picture isn’t one-directional. Anthropic still holds a wide lead on SWE-Bench Pro, where Fable 5’s 80.3% sits far above the 64.6% OpenAI reports for Sol — a roughly 15-point gap on one of the most closely watched agentic-coding benchmarks.
Independent observers have pointed out that Terminal-Bench and SWE-Bench Pro reward different kinds of work: terminal automation favours Sol’s Codex-style harness, while whole-codebase issue resolution still favours Claude.
To understand why the two companies can each claim a win, it helps to think of the two tests as different kinds of job interviews for the same candidate.
SWE-Bench Pro is like handing someone a half-finished house that a previous contractor left behind, and asking them to find exactly what’s wrong and fix only that, without touching anything else or causing a different room to leak.
It rewards a model that can read carefully, navigate something large and unfamiliar, and make a precise, contained repair — the way a software engineer fixes a specific bug in a codebase they didn’t build.
Terminal-Bench 2.1, on the other hand, is less about fixing one specific thing and more about general skilfulness around the whole house. One has to evaluate whether this person can set up the electricity, get the plumbing running, and recover if a step goes wrong midway.
It’s a broader test of operating a computer’s control panel that includes installing software, configuring servers, running multi-step jobs, and troubleshooting failures along the way.
A model can be a meticulous bug-fixer without being a great all-around operator, and vice versa. This is roughly why Fable 5 leads on the “fix this one thing precisely” test while GPT-5.6 Sol leads on the “handle this whole messy workflow” test.
There’s a second wrinkle worth knowing as these scores aren’t just testing the AI model itself but they’re also testing the toolkit it’s given to work with.
OpenAI has evaluated its models using its own in-house “Codex” toolkit, tuned to work well with how its models like to operate, a bit like letting the operator using their own tools versus renting them a standard set of equipment. A contractor will be adept at executing well with the tools he is used to, but may find it difficult to use a different toolkit.
In a normal scenario, this contractor will require some additional time to get used to the new tools before they start working. And even when they work with the new tools, it will take some time to gain tacit knowledge. If one were to factor this into GPT 5.6 Sol’s capability, then its lead with Fable 5 will be narrower.
A broader product fight
Apart from the GPT-5.6 family of models, the company also announced ChatGPT Work, an agentic mode built on GPT-5.6 that can read emails, summarise texts, create presentations, and organise inchoate ideas into charts and dashboards through a new “Sites” feature, as well as run scheduled or recurring tasks.
This product is a direct competitor to Claude Cowork, which can similarly complete multi-step, tool-using tasks for users, though the two differ in their plugin ecosystems and how deeply each is woven into its maker’s coding tools — Codex for OpenAI, Claude Code for Anthropic.
So how do the two flagship frontier labs’ models compare and contrast? Based on what the benchmark and eval scores show, both are strongly capable and land close to each other overall. Each comes with a different specialty rather than one model being categorically ahead.
The place where OpenAI clearly shows distinction is in token efficiency. On the Artificial Analysis Coding Agent Index, Sol matches or exceeds Fable 5’s score while using less than half the output tokens, taking less than half the time, and costing about one-third less.
On the broader Artificial Analysis Intelligence Index, Sol comes within a single point of Fable 5 while finishing tasks in 61% less time at roughly half the cost.
For teams weighing cost per completed task rather than raw benchmark scores this efficiency gap may be a deciding factor than which model tops any individual leaderboard. Though, as with the Codex-versus-neutral-harness comparison above, it’s worth checking whether that efficiency holds up outside OpenAI’s own tuned environment before betting a production workflow on it.
Published – July 10, 2026 04:13 pm IST
