GPT-6 Astra on OpenRouter

Original Article ↗ Hacker News Discussion ↗

OpenAI’s new GPT‑6 Astra model, now accessible via OpenRouter, Codex, and some ChatGPT tiers, is being praised for major gains in reasoning, SVG generation, and complex web/UI reconstruction—but at a steep per-token cost. Commenters weigh whether Astra’s higher quality and apparent token efficiency justify its price compared to GPT‑5.6 Sol, Luna, and cheaper Chinese models, especially for businesses sensitive to “token-maxing” behavior. There is also scrutiny of Azure’s more expensive, zero‑data‑retention offering, rate limits, and reliability issues with third‑party gateways like OpenRouter, underscoring how deployment choices hinge as much on trust and governance as on raw capability.

Access and rollout

  • Astra became available to Pro/Plus and Business users, though timing varied by account and region (e.g., Codex-only access for some Europeans, no ChatGPT).
  • Some users received “bankable resets” for each day they lacked Astra access and used them to test Sol in the meantime.
  • Astra is also exposed via OpenRouter and Azure; some report early “Not Found” issues on OpenRouter and Copilot/Foundry tool-call restrictions when reasoning is enabled.

Azure, ZDR, and enterprise concerns

  • Confusion over whether Astra on Azure is truly Zero Data Retention (ZDR) and what “privacy-preserving abuse monitoring” entails.
  • Complaints that Azure’s UI makes it hard to distinguish ZDR vs non‑ZDR models or to block non‑ZDR options.
  • Azure is said to be more expensive but offers ZDR guarantees and Azure identity integration; unclear if Codex usage is ZDR.

Capabilities: SVG, coding, and vision

  • Astra is widely praised for SVG generation: “pelican on a bicycle” comparisons show substantially higher visual quality and consistency vs Luna, Sol, Terra and older models.
  • Users debate whether this benchmark is now “overfitted” or still useful; some suspect explicit training, others attribute it to general SVG improvements.
  • Additional SVG tests (ping‑pong hamster, other animals on scooters) show strong stylistic output but clear geometric and anatomical mistakes.
  • Astra’s vision + web-dev combo is highlighted: it can closely replicate complex, flowing SVG-based designs from a static mockup, outperforming other models in layout fidelity.
  • In coding scenarios, Astra has quickly found subtle bugs (e.g., resource lifetimes) and produced solid Blender and web code, though at least one report calls its Odin code very poor.

Speed, tokens, and pricing

  • Astra often “feels faster” despite lower tokens-per-second, likely due to fewer internal “thinking” tokens.
  • Multiple users emphasize that Astra uses fewer total tokens per task vs Sol/Opus in some benchmarks; others cite metrics where Sol high/medium matches Astra low in “intelligence” at much lower $/M tokens.
  • Overall: strong disagreement on whether Astra is cheaper per task; consensus that per‑token price is high.
  • Example: a single simple frontend build via Astra cost ~$24 in API usage, seen as trivial by some (for final designs) and non‑trivial by others (limited room for experimentation).
  • Some organizations are limiting access after seeing large bills, especially from “token maxers.”

Model comparisons and orchestration

  • Users commonly use Luna for everyday/cheap work, Terra or Sol for more complex tasks, and reserve Astra/Fable for hardest problems.
  • One approach: use a “frontier” model (Sol/Astra) as an orchestrator supervising a cheaper Luna sub‑agent that does bulk reading/implementation, reducing frontier token usage.
  • Discussion of whether Astra is better than Sol at the same “effort” level: some say Astra medium/high is strictly better or equal at similar or lower per‑task cost; others dispute based on external benchmark data.

Benchmarks and evaluation debates

  • Pelican SVG gallery and cost grids are praised as intuitive model-comparison tools.
  • Concerns that repeatedly using the same pelican prompt encourages vendors to optimize specifically for it, reducing its value as a general benchmark.
  • Alternative prompts (lemur/zebra on scooters, non‑standard viewpoints) still expose weaknesses, e.g., occlusion, anatomy, and consistency.

Limitations and failure modes

  • SVG outputs, while greatly improved, still show incorrect perspectives, extra or missing limbs/mouths, malformed tables, and wrong positions in ping‑pong scenes.
  • Some note frequent lack of helmets, anatomically odd pelican knees, and incorrect bicycle details (e.g., spoke patterns, missing chains) as examples of incomplete “understanding.”
  • Tool-call failure rates on OpenAI’s Astra reportedly higher than Azure’s in one dataset, though details are sparse.

User experience and ecosystem issues

  • A few users report getting Astra access without extra resets; others accumulate multiple.
  • One user reports being auto-suspended on OpenRouter after depositing funds, with no effective support and “no refunds” stance, prompting warnings and suggestions to pursue chargebacks.
  • Skepticism toward “AGI” marketing is common; some feel media and “normies” are overhyping capabilities, given evident basic errors.
  • Broader pricing landscape: some compare Astra’s $10/$50 tiers unfavorably to much cheaper Chinese models, predicting adoption challenges unless cost-per-task clearly wins.