The frontier LLM
for coding.
State-of-the-art agentic coding — built for real engineering work, from quick edits to multi-file changes.
Available on OpenRouter. Private by design. Closed beta.
A8E 1.0
Five tiers, one family. Available via OpenRouter — from high-throughput quick edits to frontier-tier results.
a8e/a8e-1.0-liteFastest, cheapest — quick edits & high-volume tasks
↑ beats Claude Haiku 4.5 · same class, higher quality / faster / cheaper
a8e/a8e-1.0-flashLow-latency default for everyday coding
↑ beats Claude Sonnet 4.6 · same class, higher quality / faster / cheaper
a8e/a8e-1.0-proBalanced — strong agentic coding
↑ beats GLM 5.2 · same class, higher quality / faster / cheaper
a8e/a8e-1.0-maxTop capability for hard, multi-file tasks
↑ beats Claude Opus 4.8 · same class, higher quality / faster / cheaper
a8e/a8e-1.0-ultraFrontier tier — top results on the hardest agentic tasks
↑ beats Claude Fable 5 · same class, higher quality / faster / cheaper
Available on OpenRouter. Access all tiers through the OpenRouter API with your existing key. Access to A8E models requires approval — request above.
Benchmarks
Every a8e tier beats its class rival — quality, speed, and cost. Fusion wins all three.
| Category · Benchmark | a8e-1.0-ultra | Claude Fable 5 | a8e-1.0-max | a8e-1.0-pro | Claude Opus 4.8 | a8e-1.0-flash | a8e-1.0-lite | GPT 5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|---|---|---|---|
| Agentic codingSWE-Bench Pro | 82.1% | 80.3% | 76% | 72.5% | 69.2% | 66% | 60.5% | 58.6% | 54.2% |
| Agentic codingTerminal-Bench 2.1 | 89.2% | 88% | 86% | 84.5% | 82.7% | 82% | 78% | 83.4% | 70.7% |
| Agentic codingFrontierCode (Diamond) | 31% | 29.3% | 22% | 18% | 13.4% | 12% | 8% | 5.7% | — |
| Knowledge workGDPval-AA (rating) | 1962 | 1932 | 1915 | 1900 | 1890 | 1850 | 1790 | 1769 | 1314 |
| Knowledge work visionGDP.pdf | 31.4% | 29.8% | 27% | 25.5% | 22.5% | 24% | 21% | 24.9% | 16.7% |
| Spatial reasoningBlueprint-Bench 2 | 41% | 38.6% | 37% | 35% | 14.5% | 33% | 29% | 36.2% | 26.5% |
| Tool useAutomationBench | 19.8% | 17.4% | 17% | 16.2% | 15.5% | 15% | 13% | 12.9% | 9.6% |
| Computer useOSWorld-Verified | 86.5% | 85% | 84.5% | 84% | 83.4% | 82.5% | 80% | 78.7% | 76.2% |
| ReasoningHumanity's Last Exam | 66.8% | 64.5% | 61% | 59% | 57.9% | 55% | 52.5% | 52.2% | 51.4% |
| CybersecurityExploitBench | 80.5% | 78% | 72% | 68% | 40% | 55% | 44% | 34% | — |
| LegalLegal Agent Benchmark | 15.6% | 13.3% | 12% | 11% | 10.4% | 8% | 5% | 2.1% | 0% |
| HealthHealthBench Professional | 67.5% | 66% | 62% | 60% | 56.9% | 55% | 52% | 51.8% | — |
| EfficiencySpeed · p95 latency | 24.3s | 44.1s | 20s | 17s | 48s | 14s | 11s | 41s | 38s |
| EfficiencyCost · per task | $0.07 | $0.34 | $0.055 | $0.045 | $0.31 | $0.035 | $0.025 | $0.28 | $0.22 |
Projected figures. a8e figures are projected internal targets, not yet independently verified. Frontier model figures are vendor-reported. Columns are ordered by quality rank; speed = p95 latency, cost = $/task (lower is better). Methodology and full results coming at GA.
Built for responsible scale
Token-cost optimization and data privacy are first-class — not afterthoughts.
Dramatically lower cost
Get frontier-tier results at a fraction of the token cost — recurring agent loops stay affordable at scale.
Efficient by default
Optimized end-to-end so you pay for results, not waste — up to 60–90% lower token cost on typical dev workloads.
Private by design
Closed beta. Your code, context, and queries are never used to train or fine-tune any public model — ever. Enterprise-grade data isolation.
Closed access
A8E is not a public product. Access is granted by invitation only, ensuring quality, safety, and honest evaluation.
Target: 60–90% lower token cost on typical dev workloads versus direct model access — with zero privacy trade-off.
Request access
A8E is invitation-only. Leave your email and we'll reach out when we open your spot.
No spam. No public product. Access is evaluated on a case-by-case basis.
Common questions
Is A8E publicly available?
No. A8E is a closed beta product. Access is granted by invitation only. Submit your email on this page and we'll review your request.
Will my code be used to train public models?
Never. A8E is private by design. Your code, queries, and context are never used to train or fine-tune any public model.
What are the benchmark results based on?
a8e figures are projected internal targets, not yet independently verified. Frontier model figures are vendor-reported. We label them honestly and will publish full methodology and independently verified results at GA.
How much can I save on token cost?
We target 60–90% lower token cost on typical dev workloads versus direct model access — with zero privacy trade-off.
How do I access A8E?
All tiers are available through the OpenRouter API with your existing key. Access requires approval — request access on this page and we'll review.
When is GA?
We don't have a public timeline. Join the waitlist to stay informed — access is evaluated on a case-by-case basis during the closed beta.