Flower Labs says Endeavor 1.1 tops GPT-5.6 Sol, using four benchmarks where top models already score near the maximum

Flower Labs, a 2023 Cambridge spinout that pairs its models with federated learning (training on sensitive data without pooling it on a central server), released Endeavor 1.1 on November 3, two months after it shipped the first version of what it calls a frontier model. The headline figure in the launch post is a composite "intelligence" score of 97.18, which Flower ranks 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 which report you read, now makes the highest-scoring general model available. The number can't carry that meaning. The reasons are worth spelling out, because launch posts across the industry build their scores the same way.

What Flower actually reported

The composite gives equal weight to four benchmarks: 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's scores on those are 94.4, 99.9, 99.9 and 94.5. Their average is exactly the advertised 97.18. Flower also reports 81.12 on MMMU-Pro, a multimodal reasoning test, but leaves it out of the composite. Image input is new in this version.

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, while AIME stayed at 99.9. Flower describes the GPQA change as about 30% fewer errors. The arithmetic holds: the error rate fell from 8.0% to 5.6%. The company also claims better results on cybersecurity analysis and multi-step agent work, but gives no figures.

Flower's model documentation states the important caveats itself. It says the Endeavor 1.0 scores reflect Flower's internal evaluation of the launch configuration; the 1.1 launch post names no outside evaluator, so those figures should be treated as company-reported too. Competitor scores are taken from vendor and third-party reports run with different harnesses and settings. Readers are told to treat cross-model comparisons as "indicative." Neither the documentation nor the launch post gives shot counts or decoding and sampling settings. Those details matter at the top of the scale. A single AIME paper has 15 questions and HumanEval has 164 problems, so a score of 99.9 on either implies an average over many repeated attempts. That is a legitimate practice, but the result then depends on exactly how the runs were done.

Why this composite cannot rank frontier models

The bigger problem is the choice of tests. LMMarketCap's benchmark-saturation tracker, refreshed on November 4, labels HumanEval "saturated": its top five models fall within about a point of each other. 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. Its reasoning is that frontier models now score 95-99% on these contests and that the top five sit within 2 points of each other. The math leg 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 that is an older paper set than the AIME 2026 questions in Flower's composite.

Average tests on which the leaders sit within a point or two of one another, and the composite will squeeze 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 prompting, sampling and harness choices alone, and Flower acknowledges that its numbers and its rivals' were produced under different conditions. A composite like this can show that a model belongs among the leaders on well-known tests. It cannot show which of those leaders 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. Flower's launch post does not say who ran its evaluations or under what settings.

What Endeavor is built from

Flower's documentation says Endeavor "builds on" leading open-weight models. Flower adds its own work on top, including UK-specific knowledge from its earlier Lizzy 7B model, through what it lists as continual pre-training, targeted post-training and "model integration." It names no base model and gives no parameter count. 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, with Flower layering on its own Lizzy reasoning model and agent-harness software.

If that account is right, matching the leaders on these benchmarks says less about original model training and more about how well Flower fine-tunes and packages other companies' models. That is real engineering, but it is different 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 model artifacts, with formats settled during onboarding. 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 inference on GPUs located in the EU, and UK-only and other regional clusters are planned. Customers can use Flower's managed hosting, a private cloud or their own servers. They also get readable reasoning traces and keep the outputs the model produces, which Flower says they may use to train their own models. At the September launch, co-founder and chief scientist Nicholas Lane put the argument plainly in remarks to the Times, 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, and 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 in a way it can't check a self-reported composite. The organizations Flower is targeting, 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: results from an independent evaluator on harder agentic or long-context tests, disclosure of which open models Endeavor is built on, and publication of the full evaluation harness. Until then, the claim the evidence supports is narrower. 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.