Supernova @SupernovaLabsAI

Is a faster lane worth the money? Same prompt, one-file Contra-style game, median of 3 runs: glm-5.3-fast: 68 s, 31k tokens, $0.276 glm-5.3: 371 s, 75k tokens, $0.148 5.4x faster for 1.9x the cost per task. Interactive loops: fast lane. Background jobs: standard. x.com/SupernovaLabsA…

Supernova @SupernovaLabsAI

a week ago

$0.28 and 68 seconds. One API call → a playable Contra clone. Loads, auto-plays 60 s, zero edits needed. GLM-5.3 on Supernova. Median of 3 runs. supernova-labs.ai

1 0 0 64 0

Our GLM-5.3 runs in FP4. Does that cost accuracy? Artificial Analysis now reruns BFCL v4, HLE and AA-LCR subsets on every provider endpoint and scores them against a reference. Both Supernova lanes land within reference. Standard lane: $0.41 per 1M tokens blended. artificialanalysis.ai/models/glm-5-3…

GLM-5.3 averaged 17.2% of daily spend on Vercel AI Gateway, Oct 19 to 25. That is more than every other open-weight model on the board, and 2x the next one. On Supernova, GLM-5.3 is $0.63 in / $1.98 out per 1M tokens. Data: Vercel AI Gateway Leaderboard, CC BY 4.0 vercel.com/ai-gateway/lea…

“Agentic inference isn’t just about generating tokens faster. Long sessions reuse context, many agents burst together, and teams need QoS without downtime. The llm-d 1.0 release brings session-aware routing, KV-cache offload, prefill/decode disaggregation and multi-tenant controls into the production serving conversation.” newsroom.ibm.com/2026-10-20-llm…

What would you try if you worried less about the token bill? We serve GLM-5.3 through an affordable API. Give that ambitious workflow a try on Supernova. supernova-labs.ai

“We’re not ambitious enough” with AI. So give your agent crash reports, repo access, and tests. Ask it to investigate recurring failures and verify its fixes.

Felix Rieseberg @felixrieseberg

2 weeks ago

Hi, I'm Felix. I'm an eng lead working on Claude. I love watching people take flight and do things they never thought possible. We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial: felixrieseberg.com/how-to-be-more…

412 3K 24K 5.4M 52K

Your harness stays. Only the base URL changes. Claude Code on GLM-5.3 takes three env vars: ANTHROPIC_BASE_URL=api.supernova-labs.ai/glm ANTHROPIC_AUTH_TOKEN= ANTHROPIC_MODEL=glm-5.3 $0.63 in / $1.98 out per 1M tokens, 55% below Z.ai list. supernova-labs.ai/developers

Try it yourself: curl api.supernova-labs.ai/glm/v1/chat/co… \ -H "Authorization: Bearer YOUR_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"glm-5.3-fast","messages":[{"role":"user","content":"your prompt"}]}' Switch to "glm-5.3" for the cost-optimized lane. Same endpoint, same key. Lowest blended price on Artificial Analysis: $0.41/M tokens. Both at 100% Endpoint Accuracy.

What we're not hiding: 76–83% of output tokens were reasoning. You're paying for thinking, not just code. In 1 of 3 runs the fast model's game hit GAME OVER 3× during the 60 s check (auto-restarts). Standard: 0 across all 3. Token count varied a lot — fast 29k–47k, standard 53k–75k. Your workload will look different.

$0.28 and 68 seconds. One API call → a playable Contra clone. Loads, auto-plays 60 s, zero edits needed. GLM-5.3 on Supernova. Median of 3 runs. supernova-labs.ai

Prompt: “Write a single-file HTML page: a glowing particle galaxy on a full-window canvas that swirls around the mouse pointer, colors shifting over time, with a small title in the corner. Output only the HTML code.” Try it: supernova-labs.ai/developers/

glm-5.3-fast wrote this from a one-line prompt in 2.9s. opened the file as it came out and just moved the cursor around.

Multi-turn agents can waste GPU time rebuilding context. vLLM's bi-directional KV transfer lets prefill reuse cache retained by decode across turns. A useful reminder: agent latency depends on context reuse as well as decode speed. vllm.ai/blog/2026-10-1…

One interesting area to watch over is multimodal decision model and how it can be applied into the realtime interactive entertainment, gaming and human interactions! Imagine what would happen if such multimodal decision model together with Video-to-Video and RLVR on ads conversion rate to game audience into ads. So much to expect! ;)

We are cooking a decision model as well ;) Excited to sharing with you guys! A quick intro: Supernova is #1 on Artificial Analysis for GLM-5.3, and we'd love to bring it to OpenRouter! I'm Luffy, the founder. Out of 26 providers, we top at Speed and Lowest Price benchmark. We hope to bring top-tier quality token service to your platform, happy to go through your quality + functionality tests whenever works. 🫡🤠

Your team is absolutely cooking! 🔥 Supernova is #1 on Artificial Analysis for GLM-5.3 (artificialanalysis.ai/models/glm-5-3…), and we'd love to bring it to OpenRouter! We'd love to give OenRouter users 7 days of free GLM-5.3, on us and exclusive to OpenRouter, to help bring new users your way. 🫡🫡 If interested, let’s chat in DM!

@thdxr @opencode fantastic work! open source is eating the world 🫡

GLM-5.3 on Artificial Analysis, two lanes. glm-5.3: lowest blended price of 26 providers, $0.41 per 1M tokens. glm-5.3-fast: #1 in output speed, 554.8 tokens/s. GLM-5.3 (max), 10k input, 72-hour median, FP4, Oct 1, 2026. OpenAI- and Anthropic-compatible. supernova-labs.ai/pricing