A UK startup says its AI model beats OpenAI's, on tests where every top model scores near perfect

Flower Labs, a 2023 Cambridge spinout, released Endeavor 1.1 on November 3, two months after the first version of what it calls a frontier model. Flower pairs its models with federated learning, which trains on sensitive data without pooling it on a central server. The launch post headlines a composite "intelligence" score of 97.18, an average of four tests. Flower ranks it ahead of OpenAI's GPT-5.6 Sol (96.25), Anthropic's Claude Fable 5 (95.30), Moonshot AI's Kimi K3 (94.83) and Nvidia's Nemotron 3 Ultra (92.28).

Read literally, that would mean a company that has raised somewhere between about £18 million and £23 million, depending on the report, now makes the highest-scoring general model available. The number can't carry that meaning. The reasons matter beyond Flower, because launch posts across the industry build their scores the same way.

What Flower actually reported

The composite weights four benchmarks, or standard tests, equally: GPQA (graduate-level science multiple choice), AIME 2026 (competition math), HumanEval (short Python programming tasks) and IFEval (following formatting and content instructions). Endeavor 1.1 scores 94.4, 99.9, 99.9 and 94.5, which average exactly the advertised 97.18. Flower also reports 81.12 on MMMU-Pro, a test of reasoning over images and text, but leaves it out of the composite.

The gains over Endeavor 1.0 are small, mainly because the tests leave little room to improve. GPQA rose from 92.0 to 94.4, HumanEval from 98.2 to 99.9 and IFEval from 94.1 to 94.5; AIME stayed at 99.9. Flower also claims better results on cybersecurity analysis and multi-step agent work, but gives no figures.

Flower's own documentation flags the key caveats. It says the Endeavor 1.0 scores come from Flower's internal evaluation. The 1.1 launch post names no outside evaluator, so those figures should be treated as company-reported too. Competitor scores come from vendor and third-party reports run with different test setups and settings. Readers are told to treat cross-model comparisons as "indicative." Neither document gives detailed test settings, such as how many worked examples the model saw or how its answers were generated. Those settings matter near the top of the scale. An AIME paper has 15 questions and HumanEval has 164 problems, so 99.9 on either implies averaging many repeated attempts. That is legitimate, but results then depend on exactly how the runs were done.

Why this score cannot rank the top models

The bigger problem is the choice of tests. LMMarketCap's benchmark-saturation tracker, refreshed on November 4, labels HumanEval "saturated," meaning it no longer separates the leaders: the top five fall within about a point. GPQA Diamond and IFEval are rated "weak," with top-five spreads of 0.7 and 1.0 points. BenchLM, a separate aggregator, stopped counting AIME and HMMT toward its math score this year, saying frontier models now score 95-99% on these contests, with the top five within 2 points. Math is the least settled of the four. LMMarketCap still rates the older AIME 2024 set as active, with a 7.5-point gap between first and fifth, though Flower's composite uses the newer AIME 2026 questions.

Average tests where the leaders sit within a point or two, and the composite squeezes them just as tightly. Endeavor's lead over GPT-5.6 Sol is 0.93 points. A gap that small could easily come from differences in test setup alone. Such a composite can show a model belongs among the leaders on well-known tests, but not which leader is best.

Outside verification is also missing. As of Wednesday, Endeavor 1.1 was available only by request, and we could find no published results for it from an independent evaluator.

What Endeavor is built from

Flower's documentation says Endeavor "builds on" leading open-weight models, whose files their makers publish for others to adapt. Flower adds its own work, including UK-specific knowledge from its earlier Lizzy 7B model. It lists the methods as continual pre-training, targeted post-training (further training of an existing model) and "model integration." It names neither the base model nor its parameter count, the usual measure of size. A Resultsense report on the September launch, which credited Business Matters and cited no further sourcing, said openly available Meta and Mistral models provide much of Endeavor's underlying capability. In that account, Flower layers on its own Lizzy reasoning model and software for running the model as an agent.

If that is right, matching the leaders on these benchmarks says less about original model training than about how well Flower fine-tunes (adapts through extra training) and packages other companies' models. That is real engineering, but it differs from what the big labs do when they train their own frontier systems. Endeavor itself is not open-weight either: the documentation lists only "private distribution" of the model files. Resultsense described the product as closed and licensed, and called that confusing for a pitch built on sovereignty.

The real pitch is control, not leaderboard rank

Most of Flower's commercial argument isn't about scores. Version 1.1 adds the option of running the model on GPUs located in the EU, with UK-only and other regional clusters planned. Customers can use Flower's managed hosting, a private cloud or their own servers. They get readable reasoning traces, a record of how the model reached its answers. They also keep the model's outputs, which Flower says they may use to train their own models. Co-founder and chief scientist Nicholas Lane put it plainly to the Times at the September launch, as quoted by Dealroom: "Europe should not have to rent its intelligence indefinitely from a handful of US companies." Dealroom's report names the NHS among Flower's clients. Flower told Resultsense its tooling is used at more than 2,500 organizations, including JP Morgan.

A buyer can check those deployment and data terms during a pilot; it can't check a self-reported composite. Flower's target customers, which Resultsense describes as hospitals, banks, defense contractors and governments, will run their own evaluations on their own workloads before committing. Leading with a leaderboard claim invites a comparison the product doesn't need to win and the evidence doesn't yet support.

Three things would change that picture. First, results from an independent evaluator on harder tests of multi-step autonomous work or very long inputs. Second, disclosure of which open models Endeavor is built on. Third, publication of the full evaluation setup. Until then, the evidence supports a narrower claim. Endeavor 1.1 is a fine-tuned system that scores near the maximum on four familiar benchmarks, sold with deployment and data terms the large US labs don't offer.