Anthropic and OpenAI are locked in an more and more tight race for frontier AI supremacy, and the hole between their flagship fashions are narrowing to a benchmark-by-benchmark contest.
Claude has led on real-world software-engineering duties with Fable 5 scoring 80.3% on SWE-Bench Professional at its June launch, properly forward of GPT-5.5’s 58.6%. However OpenAI’s newly launched GPT-5.6 Sol claims the state-of-the-art spot on the Synthetic Evaluation Coding Agent Index at 80, narrowly forward of Fable 5, whereas utilizing lower than half the tokens and time.
Since launching Mythos and giving a choose group of corporations in its Glasswing Alliance early entry, Anthropic has positioned itself on the entrance of the frontier race. It adopted up by releasing Fable 5, a model of Mythos with further safeguards round biology, cybersecurity, and AI analysis, to a much wider set of customers, cementing that lead.
That momentum hit a snag when a dispute with the U.S. authorities over export controls led Anthropic to be handled as a supply-chain danger, forcing it to droop entry to Fable 5 and Mythos 5 for 19 days — from June 12 to July 1, 2026 — earlier than the Commerce Division lifted the related controls and entry was restored.
Simply over per week after Fable 5 got here again on-line for all customers, OpenAI unveiled its personal next-generation household: GPT-5.6 Sol, Terra, and Luna. Somewhat than a single flagship, OpenAI break up the discharge into three functionality tiers priced individually with Sol on the prime and Luna on the backside, arguing this lets every tier advance by itself tempo.
OpenAI has been vocal that Sol edges out Anthropic’s flagship on a number of fronts. On Brokers’ Final Examination, an analysis of long-running skilled workflows throughout 55 fields, Sol scored 53.6, which OpenAI says is 13.1 factors forward of Fable 5. On the Synthetic Evaluation Coding Agent Index, Sol posted a brand new excessive of 80, about 2.8 factors above Fable 5. It additionally claims new state-of-the-art marks on BrowseComp (92.2%) and OSWorld 2.0 (62.6%).
Image isn’t one-directional
However the image isn’t one-directional. Anthropic nonetheless holds a large lead on SWE-Bench Professional, the place Fable 5’s 80.3% sits far above the 64.6% OpenAI stories for Sol — a roughly 15-point hole on one of the crucial intently watched agentic-coding benchmarks.
Unbiased observers have identified that Terminal-Bench and SWE-Bench Professional reward totally different sorts of labor: terminal automation favours Sol’s Codex-style harness, whereas whole-codebase situation decision nonetheless favours Claude.
To grasp why the 2 corporations can every declare a win, it helps to think about the 2 assessments as totally different sorts of job interviews for a similar candidate.
SWE-Bench Professional is like handing somebody a half-finished home {that a} earlier contractor left behind, and asking them to seek out precisely what’s incorrect and repair solely that, with out touching the rest or inflicting a unique room to leak.
It rewards a mannequin that may learn rigorously, navigate one thing giant and unfamiliar, and make a exact, contained restore — the best way a software program engineer fixes a particular bug in a codebase they didn’t construct.
Terminal-Bench 2.1, alternatively, is much less about fixing one particular factor and extra about normal skilfulness round the entire home. One has to judge whether or not this individual can arrange the electrical energy, get the plumbing working, and get better if a step goes incorrect halfway.
It’s a broader check of working a pc’s management panel that features putting in software program, configuring servers, working multi-step jobs, and troubleshooting failures alongside the best way.
A mannequin generally is a meticulous bug-fixer with out being an important all-around operator, and vice versa. That is roughly why Fable 5 leads on the “repair this one factor exactly” check whereas GPT-5.6 Sol leads on the “deal with this entire messy workflow” check.
There’s a second wrinkle price realizing as these scores aren’t simply testing the AI mannequin itself however they’re additionally testing the toolkit it’s given to work with.
OpenAI has evaluated its fashions utilizing its personal in-house “Codex” toolkit, tuned to work properly with how its fashions prefer to function, a bit like letting the operator utilizing their very own instruments versus renting them an ordinary set of kit. A contractor might be adept at executing properly with the instruments he’s used to, however might discover it tough to make use of a unique toolkit.
In a standard situation, this contractor would require some further time to get used to the brand new instruments earlier than they begin working. And even once they work with the brand new instruments, it can take a while to realize tacit data. If one have been to issue this into GPT 5.6 Sol’s functionality, then its lead with Fable 5 might be narrower.
A broader product battle
Aside from the GPT-5.6 household of fashions, the corporate additionally introduced ChatGPT Work, an agentic mode constructed on GPT-5.6 that may learn emails, summarise texts, create displays, and organise inchoate concepts into charts and dashboards via a brand new “Websites” characteristic, in addition to run scheduled or recurring duties.
This product is a direct competitor to Claude Cowork, which might equally full multi-step, tool-using duties for customers, although the 2 differ of their plugin ecosystems and the way deeply every is woven into its maker’s coding instruments — Codex for OpenAI, Claude Code for Anthropic.
So how do the 2 flagship frontier labs’ fashions evaluate and distinction? Primarily based on what the benchmark and eval scores present, each are strongly succesful and land shut to one another total. Every comes with a unique specialty slightly than one mannequin being categorically forward.
The place the place OpenAI clearly reveals distinction is in token effectivity. On the Synthetic Evaluation Coding Agent Index, Sol matches or exceeds Fable 5’s rating whereas utilizing lower than half the output tokens, taking lower than half the time, and costing about one-third much less.
On the broader Synthetic Evaluation Intelligence Index, Sol comes inside a single level of Fable 5 whereas ending duties in 61% much less time at roughly half the fee.
For groups weighing price per accomplished activity slightly than uncooked benchmark scores this effectivity hole could also be a deciding issue than which mannequin tops any particular person leaderboard. Although, as with the Codex-versus-neutral-harness comparability above, it’s price checking whether or not that effectivity holds up outdoors OpenAI’s personal tuned atmosphere earlier than betting a manufacturing workflow on it.
Printed – July 10, 2026 04:13 pm IST