Endeavor 1.1: Flower Labs adds EU inference and image inputs, claiming the top composite score over OpenAI and Anthropic

Flower Labs has updated its frontier model with the two things regulated buyers keep asking for: somewhere in Europe to run it, and a way to see what it was thinking. The company introduced Endeavor 1.1 on November 3, 2026, its third model release this year after Lizzy and Endeavor 1.0, with stronger reasoning for demanding agent work, image inputs alongside text, and inference on GPUs located in the European Union. The pitch is not just a better score. It is that customers should not have to choose between frontier intelligence and control over where the model runs, what it does with their work, and when it changes.

The numbers Flower is leading with

Flower says Endeavor 1.1 achieves the highest score on its own intelligence composite, an equal-weight average of four benchmarks — GPQA (science), AIME 2026 (mathematics), HumanEval (coding) and IFEval (instruction following) — at 97.18, ahead of GPT-5.6 Sol at 96.25 and Claude Fable 5 at 95.30, with Kimi K3 at 94.83 and Nemotron 3 Ultra at 92.28. The composite is Flower’s construction, built from benchmarks it selected and weighted equally, so read it as the company’s account of where its model stands, not as an independent ranking.

The version-on-version gains Flower reports are specific: GPQA rises from 92.0 to 94.4, which Flower says represents about 30% fewer errors; HumanEval rises from 98.2 to 99.9; IFEval moves from 94.1 to 94.5; AIME 2026 stays at 99.9. A fifth benchmark, MMMU-Pro, is reported separately at 81.12 and is not part of the composite. Flower also says it is seeing similar gains in cybersecurity analysis and complex agent workflows, without publishing scores for those. None of these figures has been independently reproduced in public, and the composite’s equal weighting is itself a choice: a different mix of tests would tell a different story.

Control is the pitch

The rest of the announcement is about ownership. Customers can inspect readable thinking traces to review the model’s approach, investigate decisions and find where it went wrong — where, Flower says, OpenAI exposes only summaries of thinking traces and Anthropic typically withholds raw traces. Outputs and traces belong to the customer, who can retain, organise and reuse them, including using them to fine-tune or train their own models, such as smaller specialist models for recurring tasks. An engineering team, Flower suggests, could collect verified fixes and their traces and train a troubleshooting model around its own systems, turning spending into intelligence that stays with the buyer.

Deployment follows the same logic. Endeavor 1.1 is available through Flower’s managed service, in a customer’s private cloud, or on premises, and a customer can start managed and move to private deployment later. Running on premises or in a private cloud keeps the model and sensitive workloads inside the customer’s environment, Flower says, so prompts that reveal proprietary code, product plans or problem-solving methods are not exposed to another company’s security or access controls. The model documentation lists a maximum context length of one million tokens, according to S3DAT’s summary of the release, with effective limits depending on deployment.

Why EU inference matters

The EU option is the most concrete change. Customers can choose Flower’s global inference offering or inference on GPUs located specifically in the EU, with service-level agreements and support from Flower. Flower argues global APIs do not guarantee inference within a specific region, and points to Azure’s UK-only per-token offering as still limited to GPT-4o and GPT-4.1 mini — Flower’s example, not an independent audit of Azure’s catalogue. UK-only inference and regional options in the Middle East, Asia and Africa are planned, Flower says, pairing capability with service commitments for production workloads.

What to watch

Access is by request: Flower invites prospective customers to test Endeavor 1.1 on their own work and compare results on the tasks that matter to them, and its post names no price. That test is the one that matters. The benchmark composite is a vendor-built snapshot, the EU and deployment options are real choices with real operational consequences, and the traces-and-ownership promise is only valuable if the traces are genuinely readable and the training rights hold up in the contract. If Flower’s model performs as claimed on a buyer’s own tasks, the control pitch — EU GPUs, private deployment, inspectable reasoning, work you keep — is a combination Flower argues the US frontier labs do not offer in one package. If it does not, no amount of deployment flexibility closes the gap.