Model Release
Announcing Starlight Pulse 3
Our fastest model just got a frontier brain. Pulse 3 is rolling out to every plan today - free included - and it lands above GPT-6.1 Sol, DeepSeek V4.1 Flash, Qwen 3.8 Max, and Fable 5.1 on our evaluation suite, at a fraction of their cost per task. Only Opus 5.5 and GPT-6 Astra still score higher.
The fastest way to say it
Pulse 3 is a ground-up retrain of the Pulse line, and it's the largest single-generation jump we've ever shipped. Same snappy latency you use Pulse for - now with reasoning depth that was Core-2 territory six months ago. It plans before it writes, checks its own output in the sandbox, and stays coherent across hour-long agentic runs. Every capability in Starlight - code, 3D, images, video, sound, agents - is powered by it from today, on every plan, including free.
Pulse 3 against the October 2026 frontier
Real, published scores from vendor launch tables and independent runs (Scale SEAL, Vals AI, Artificial Analysis) where available, measured against Pulse 3's internal evaluation runs, October 2026. Pulse 3 sits above GPT-6.1 Sol, DeepSeek V4.1 Flash, Qwen 3.8 Max, and Fable 5.1 - and below Claude Opus 5.5 and GPT-6 Astra, which we're happy to call the two best models on the planet right now. Competitor names are trademarks of their respective owners; Starlight is not affiliated with any of them. Scores vary with harness and effort settings - run your own evals before you bet a codebase on any number, including ours.
SWE-bench Pro
Real GitHub issues in professional codebases (vendor aggregate, September 2026)
| Claude Opus 5.5 | 89.9 |
| Starlight Pulse 3 | 84.6 |
| Claude Fable 5.1 | 81.2 |
| Qwen 3.8 Max | 67.7 |
| GPT-5.6 Sol | 64.6 |
AutomationBench
End-to-end business workflow orchestration (Zapier leaderboard)
| GPT-6 Astra | 41.4 |
| Claude Opus 5.5 | 40.0 |
| Starlight Pulse 3 | 38.9 |
| GPT-6.1 Sol | 36.1 |
| Claude Fable 5.1 | 31.4 |
GPQA Diamond
Graduate-level science Q&A (independent runs, max effort)
| GPT-6 Astra | 96.0 |
| Starlight Pulse 3 | 95.1 |
| Claude Fable 5.1 | 93.7 |
| Qwen 3.8 Max | 92.6 |
| DeepSeek V4.1 Flash | 90.9 |
Terminal-Bench 4.0
Long-horizon terminal and agentic work (Vals AI harness, Oct 7)
| Claude Opus 5.5 | 65.2 |
| GPT-6 Astra | 59.6 |
| Starlight Pulse 3 | 58.4 |
| Claude Fable 5.1 | 58.1 |
| GPT-6.1 Sol | 55.1 |
| DeepSeek V4.1 Flash | 31.2 |
Creative Generation (composite)
3D generation + code + audio, Starlight's own evaluation suite
| Starlight Pulse 3 | 91.2 |
| Claude Opus 5.5 | 53.6 |
| GPT-6 Astra | 51.9 |
| Claude Fable 5.1 | 45.8 |
| GPT-6.1 Sol | 43.2 |
| DeepSeek V4.1 Flash | 38.7 |
| Qwen 3.8 Max | 35.4 |
How to read these numbers
We did not cherry-pick a single chart where Pulse 3 wins. On agentic coding, Opus 5.5 (89.9 SWE-bench Pro, 66.4 Terminal-Bench 4.0 at vendor-reported effort) and GPT-6 Astra (96.0 GPQA Diamond, 59.6 Terminal-Bench 4.0 on the Vals harness) are still ahead of us, and we expect them to stay ahead for a while. What Pulse 3 does is collapse the distance: it trades punches with Fable 5.1, beats GPT-6.1 Sol and Qwen 3.8 Max on reasoning and terminal work, beats DeepSeek V4.1 Flash on everything except raw throughput - and it does all of it inside Starlight's sandbox with 240+ skills, a 1M-token window, and creative generation that no frontier lab even attempts.
What actually changed in Pulse 3
The benchmark deltas are downstream of five structural upgrades.
Adaptive thinking, tuned for speed
Pulse 3 decides per-request how hard to think, then caps itself. Easy questions answer in a beat; hard ones get extended chain-of-thought without you touching an effort dial. You get Core-class depth only when the task pays for it.
Repo-scale coding
84.6 on SWE-bench Pro means Pulse 3 resolves real issues in professional codebases - multi-file diffs, hidden tests, no hints. It plans the change, writes it, runs it in the sandbox, and reads its own test output before replying.
Long-horizon endurance
The Terminal-Bench 4.0 and agent-loop upgrades are about not falling apart at step forty. Pulse 3 keeps state across hundreds of tool calls, recovers from its own mistakes, and finishes the job instead of narrating it.
Creation, not just completion
Rigged 3D meshes, textured and dropped into your scene. Audio that loops. Images that match your art direction. The creative composite score isn't a vanity number - no competitor above or below us on this page runs that pipeline natively.
1M context that's actually used
A million tokens is worthless if retrieval decays. Pulse 3 holds 90%+ needle accuracy at 128K and stays above 80% at 1M on our MRCR-v2 runs - monorepos and full-session history stay in reach.
Two models on Earth out-score Pulse 3. Neither one can rig a character mesh, score its cutscene, and ship the code that drives both - in one conversation, on the free plan.
Starlight Team
Availability
Pulse 3 is the default model on Starlight starting today, on every plan: Drift (free), Orbit, Flare, and Nova. Thinking effort scales with your plan - Light and Medium everywhere, High and Ultra on paid tiers. Core 2 remains available for the hardest reasoning work, and you can switch between them mid-conversation; context follows you. No migration needed - your existing sessions just got faster.
Try Pulse 3 now
Free on the Drift plan. No credit card, no waitlist - open a chat and give it something hard.