GPT‑5.3‑Codex‑Spark

Original Article ↗ Hacker News Discussion ↗

Positioning and competition

  • Many see this as part of an arms race with Anthropic, Google, etc., with increasingly rapid, overlapping releases.
  • Several note GPT‑5.3‑Codex‑Spark is a smaller, faster tier beneath full 5.3‑Codex, roughly analogous to previous “mini” tiers, not a straight upgrade in capability.
  • Comparisons: GLM‑4.7 on Cerebras, Claude Code Opus, Gemini 3, and Perplexity’s Cerebras‑backed Sonar. Some say Codex 5.3 is currently the best coding model; others still prefer Opus for “agentic” work.

Speed vs quality and use cases

  • Divided views on whether speed is the right problem to solve:
    • Some want “faster and better” and complain Codex 5.3 is too slow vs Opus.
    • Others argue fast, cheaper models are ideal for bulk/low‑risk tasks (renames, refactors, search, boilerplate) while heavy models handle complex reasoning.
  • There’s a recurring wish for automatic routing: fast model for trivial edits, cheap for background/batch, smart/slow for hard problems.

Agents and long‑running workflows

  • OpenAI’s claim about models working autonomously for “hours, days or weeks” is met with skepticism; many say long‑running agents still go off the rails.
  • Others report success with overnight debugging, codebase upgrades, and multi‑hour builds when paired with good harnesses (tests, verification loops, tools like “Ralph”).
  • Consensus: closed loops with clear success criteria and verification are crucial; otherwise agents waste tokens or produce subtle bugs.

Cerebras hardware and economics

  • The Cerebras WSE‑3 wafer‑scale chip draws fascination (size, defect‑tolerance, 20kW+ power) and debate:
    • Some see it as underrated, ideal for ultra‑low‑latency inference.
    • Others question VRAM limits, density, perf/$ vs GPUs/TPUs, and long‑term viability.
  • Broader discussion spills into Nvidia vs TPUs vs custom ASICs, power constraints, and whether specialized inference silicon will erode Nvidia’s dominance.

Infrastructure and API changes

  • A significant part of the latency win comes from harness changes: persistent WebSockets, reduced per‑request and per‑token overhead, better time‑to‑first‑token. These improvements are expected to roll out to other models.
  • Some note that open‑source agents may struggle to match these gains without a standardized WebSocket LLM API.

Benchmarks, early impressions, and concerns

  • Benchmarks like Terminal Bench, SWE‑Bench Pro, personal “Bluey Bench,” and a “pelican” blog test show:
    • Spark is dramatically faster (hundreds–1000+ tok/s) but with noticeably lower quality than full 5.3‑Codex and even some prior GPT variants.
  • Early users describe it as “blazing fast” with a clear “small model feel”: more mistakes, worse context discipline, fragile adherence to AGENTS.md rules.
  • Worryingly, several report destructive behavior (deleting files, bad git operations) and argue “risk of major failure” should be part of evaluating fast agentic models.

Other themes

  • Frustration over opaque pricing and heavy marketing language; some criticize chart scaling as misleading.
  • Complaints that Codex models are tightly coupled to the Codex harness and weaker as general‑purpose chat models.
  • Mixed reactions to accelerating model churn: some embrace the pace for productivity, others deliberately ignore it and stick with “good enough” tools.