Talos-1 is our production model for agentic work in European companies: reading contracts, calling internal APIs, filling in forms, finishing a job inside the systems a company already runs. It is built and served inside the EU.
- Context
- 200k tokens at launch. The model supports one million; we widen the window in phases, to 512k later this year and to one million after that.
- Reasoning
- Three efforts: low, medium and high. Default: medium
- Input
- Text, images, video and files
- Output
- Text
- Pricing
- €1 per million input tokens · €4 per million output tokens · €0.10 per million cached input tokens
- Access
- Hosted by Ablatic and delivered as an OpenAI-compatible API to customers. No self-service signup yet; it opens once the capacity is there. Companies reach out and get a reply within 24 hours.
- Data
- EU data centres, no training on customer data, Ablatic acts as GDPR processor under an Art. 28 DPA. Details on the compliance page.
Benchmarks
| Benchmark | Talos-1 | Claude Opus 5 | GPT-5.6 Sol | DeepSeek V4 Pro | Kimi K3 | GLM-5.3 | Qwen3.8 Max |
|---|---|---|---|---|---|---|---|
| Agentic | |||||||
| Agents' Last Exam (Score) | 49.0 | 55.2 | 53.6 | 27.6 | 51.6 | - | 52.5 |
| AutomationBench, public | 42.5 | 50.3 | 45.8 | 43.2 | 46.7 | 48.2 | 39.8 |
| Terminal-Bench 4.0 | 26.5 | 51.8 | 37.3 | 14.1 | 12.6 | 41.8 | 11.1 |
| Toolathlon | 74.4 | 80.6 | 74.9 | 74.4 | 76.5 | 73.0 | 72.5 |
| BrowseComp | 86.7 | 90.8 | 90.4 | 83.4 | 91.2 | - | - |
| Coding | |||||||
| DeepSWE v1.1 | 51.8 | 73.6 | 72.7 | 62.7 | 67.5 | 66.9 | 56.6 |
| NL2Repo-Bench | 47.3 | - | - | 61.5 | 58.0 | 58.0 | 55.9 |
| Vision and documents | |||||||
| OmniDocBench | 90.8 | - | 92.9 | text | 91.1 | text | 92.1 |
| MMMU-Pro | 80.9 | 84.7 | 83.4 | text | 80.5 | text | 82.4 |
| General | |||||||
| HLE without tools | 34.7 | 54.9 | 49.5 | 41.0 | 46.9 | 42.3 | 43.6 |
| HLE with tools | 49.3 | 64.7 | 64.5 | 60.0 | 56.0 | 62.5 | 56.2 |
| GPQA Diamond | 90.1 | 93.2 | 94.1 | 92.8 | 93.5 | 91.7 | 93.5 |
| IFBench | 82.2 | - | 72.7 | 76.5 | - | 73.1 | 82.8 |
Bold marks the best score in each row among the comparison models. Talos-1 ran at reasoning effort high, the comparison models at their highest effort (DeepSeek V4 Pro is the 0813 release). We measured every Talos-1 number with each benchmark's own harness. Terminal-Bench 4.0 ran under Claude Code. NL2Repo ran under OpenHands, the harness its authors use. BrowseComp and HLE with tools ran in our own harness with web search and Python, 200 interaction steps and a 512k context limit. Agents' Last Exam shows the Score metric on the Overall split of the official leaderboard; GLM-5.3 is not listed there. Scores for the other models come from the official leaderboards and from the vendors' own reports. "text": the model takes no images, so the benchmark does not apply. "-": no published score. Sources for every comparison score are listed below.
What the benchmarks measure
- Agents' Last Exam (UC Berkeley): long, real-world work tasks on a computer, terminal and screen together, scored on the finished result.
- AutomationBench (Zapier): 600 business workflows across 47 simulated SaaS tools, run end to end.
- Terminal-Bench 4.0: hard tasks in a Linux shell, from server setup to build repair.
- Toolathlon: tasks in real software such as mail, calendar, spreadsheets and GitHub, driven through the tools' own interfaces.
- BrowseComp (OpenAI): hard questions whose answers are scattered across the web and must be found by browsing.
- DeepSWE v1.1 (Datacurve): fixing real software issues inside existing code repositories.
- NL2Repo-Bench: building a whole repository from a written specification.
- OmniDocBench: reading real documents, including layouts, tables and formulas.
- MMMU-Pro: multimodal reasoning over charts, diagrams and images.
- Humanity's Last Exam: frontier-research questions across a hundred disciplines, without and with tools.
- GPQA Diamond: graduate-level science questions.
- IFBench: following written instructions exactly.
Where the comparison scores come from
Read on 14 September 2026. Comparison models ran under their vendors' own agents where the board lists one.
- Agents' Last Exam: agents-last-exam.org/leaderboard, Score metric, Overall split.
- AutomationBench: github.com/zapier/AutomationBench, README, public set.
- Terminal-Bench 4.0: tbench.ai/leaderboard.
- Toolathlon: toolathlon.xyz/docs/leaderboard; Claude Opus 5 from benchlm.ai.
- BrowseComp: OpenAI launch post (GPT-5.6 Sol, single agent), Kimi K3 model card, benchlm.ai for Claude Opus 5 and DeepSeek V4 Pro.
- DeepSWE v1.1: deepswe.datacurve.ai; GLM-5.3 and Qwen3.8 Max from the GLM-5.3 model card.
- NL2Repo: GLM-5.3 model card (huggingface.co/zai-org/GLM-5.3) and benchlm.ai.
- OmniDocBench: Kimi K3 model card (huggingface.co/moonshotai/Kimi-K3); Qwen3.8 Max from the OmniDocBench 1.5 listing on benchlm.ai.
- MMMU-Pro, GPQA Diamond, HLE without tools, IFBench: artificialanalysis.ai; IFBench for Qwen3.8 Max from benchlm.ai.
- HLE with tools: vendor model cards (GLM-5.3, Kimi K3, DeepSeek V4); Claude Opus 5 from benchlm.ai.