GPT-5.3-Codex-Spark

GPT-5.3-Codex-Spark, publicly referred to as "GPT-5.3-Codex-Spark" or "Codex-Spark," is OpenAI's distilled, low-latency sibling of GPT-5.3-Codex, launched as a research preview on 2026-02-12 (Cerebras blog; original post timestamped the same day on the OpenAI Developer Community). It runs on Cerebras' Wafer-Scale Engine 3 (WSE-3) and is the first shipped release of the OpenAI–Cerebras partnership announced the previous month, January 2026 (Cerebras blog; OpenAI Developer Community). At launch, general access is a ChatGPT Pro-only research preview inside the Codex app, Codex CLI, and VS Code extension; Cerebras' own announcement states API access is "rolling out to select design partners" only, not the general API population (Cerebras blog; Codex pricing docs). OpenAI has not published a standalone API model-ID string for Spark as of 2026-07-12.

Key specs

Values as of the 2026-02-12 launch, re-verified 2026-07-12 directly against the vendor sources below. Spark never shipped a public API SKU during this research-preview period and is absent from OpenAI's standard model-docs catalog, so several fields below have no vendor-disclosed value.

Spec	GPT-5.3-Codex-Spark	Source
Research-preview launch	2026-02-12	Cerebras blog; OpenAI Developer Community
Context window	128,000 tokens	OpenAI Developer Community
Max output tokens	Not disclosed	—
Modalities	Text input, text output only (no image input)	OpenAI Developer Community; Help Net Security
Knowledge cutoff	Not disclosed for Spark specifically	—
Throughput	Over 1,000 tok/s at launch; over 1,200 tok/s (about 30% higher) after a follow-up infrastructure update on 2026-02-20	Cerebras blog; OpenAI Developer Community
Inference hardware	Cerebras Wafer-Scale Engine 3	Cerebras blog
API access	Not generally available; limited to "select design partners"	Cerebras blog
Pricing (general access)	Not disclosed; no finalized credit rate — the rate card lists "research preview" instead of a per-Mtok figure, and usage is capped by a separate Cerebras-capacity-based limit rather than standard quota	Codex pricing docs
Access tier	ChatGPT Pro subscribers only, via Codex app / Codex CLI / VS Code extension	Codex pricing docs

The models.dev registry (snapshot as of 2026-07-12) lists a different set of values for this slug — 32,000 max output tokens, image+PDF+text input modalities, a 2026-02-05 release date, and API pricing identical to the parent model ($1.75 / $0.175 cached / $14.00 per Mtok). None of these four values could be confirmed in Spark-specific vendor material (Cerebras blog, OpenAI Developer Community thread, or Codex pricing docs); the 2026-02-05 date matches the parent GPT-5.3-Codex system card, and the pricing matches the parent model's API price, suggesting the registry entry inherited parent-model fields rather than measuring Spark directly. This page follows the vendor sources above instead.

For scale, the parent GPT-5.3-Codex (system card dated 2026-02-05) ships a 400,000-token context-window, 128,000-token max output, a 2025-08-31 knowledge cutoff, and $1.75 / $0.175 (cached) / $14.00 per-Mtok API pricing (GPT-5.3-Codex model docs, accessed 2026-07-12) — Spark trades most of that context and essentially all general API access for raw token throughput.

Benchmark results

No first-party numeric benchmark score for GPT-5.3-Codex-Spark itself has been published as of 2026-07-12. Cerebras' launch post states only, qualitatively, that Spark delivers "more capable responses than GPT-5.1-Codex-mini while completing tasks in a fraction of the time" and that it "outperform[s] GPT-5.1-Codex-mini on SWE-Bench Pro and Terminal-Bench 2.0" — a comparison against a smaller sibling model, with no numeric table in the post (Cerebras blog).

Spark has no entry on the official Scale Labs SWE-bench Pro leaderboard as of 2026-07-12. This project's SWE-bench data snapshot, sourced from the Vals AI SWE-bench Verified leaderboard as of 2026-07-12, records the parent GPT-5.3-Codex at 78.0% and explicitly notes that Spark itself does not appear on that leaderboard — there is no comparable, leaderboard-verified data point for Spark on any benchmark this project's data pipeline tracks.

A specific figure (roughly 56% SWE-bench Pro for Spark versus roughly 72% for full GPT-5.3-Codex) circulates across aggregator blogs, but none of them cite a traceable primary source, and at least one explicitly labels its own Terminal-Bench number for Spark "(estimated)." Treat any such percentage as unconfirmed until it is traced to an OpenAI chart or an official leaderboard entry. No chart is embedded on this page: neither official leaderboard lists Codex-Spark, and this project's data pipeline does not treat single-anecdote demos or unattributed aggregator figures as chartable data.

One non-benchmark, vendor-adjacent data point exists: in a live demo building a snake game, Spark finished in about 9 seconds versus roughly 43 seconds for standard GPT-5.3-Codex in the same demo (ServeTheHome) — a single anecdotal demo, not a benchmark result.

Agentic behavior notes

Spark targets interactive, in-editor workflows with a developer in the loop, not long-horizon autonomous runs: by default it makes minimal, targeted edits and does not proactively run a test cycle unless the prompt asks for one, unlike full GPT-5.3-Codex and GPT-5.5 (Help Net Security).

No dedicated system card or Preparedness Framework section for Spark specifically exists as of 2026-07-12: the word "Spark" appears in neither the GPT-5.3-Codex system card (2026-02-05) nor the official Deployment Safety Hub cyber-range page for GPT-5.3-Codex (System Card PDF; Deployment Safety Hub). The parent system card classifies GPT-5.3-Codex as High capability under the Cybersecurity domain on a precautionary basis ("we do not have definitive evidence that this model reaches our High threshold, but are taking a precautionary approach") and High risk in the Biological/Chemical domain, without reaching High capability on AI self-improvement (System Card PDF, pp. 3, 6) — none of this is confirmed to extend to Spark, since the card never mentions it.

An issue filed against the official openai/codex CLI repository during the research-preview rollout documents early tooling friction: users reported that Spark's usage limits surfaced in the CLI's /status output and, in one case, in the ChatGPT web UI for a Plus subscriber who should not have had Spark access, while /model gave no clear signal of eligibility. An OpenAI contributor replied that "the spark model is currently available only to Pro subscribers"; the issue was subsequently closed (GitHub issue #12992, accessed 2026-07-12).

Limitations

• No published benchmark scores. Neither OpenAI nor an official leaderboard has published a numeric Spark score for SWE-bench Pro or Terminal-Bench as of 2026-07-12; evaluate it on your own workload before trusting any third-party percentage in circulation.
• Smaller context window. 128K tokens versus the parent's 400K constrains use on large codebases or long agent sessions (OpenAI Developer Community).
• Text-only. No image input at launch, unlike parent GPT-5.3-Codex (OpenAI Developer Community).
• API access is not general. ChatGPT Pro research preview is the only broadly available surface; Cerebras' own announcement limits API access to "select design partners," so teams wiring Spark into an automated harness or CI pipeline should not assume they can reach it programmatically as of 2026-07-12 (Cerebras blog; Codex pricing docs).
• Thin hardware disclosure. The only public description of Spark's inference stack is Cerebras' own announcement (WSE-3, over 1,000 tok/s); neither company has published whether an alternate hardware path exists, so treat throughput and availability as coupled to this one named partner until stated otherwise (Cerebras blog).
• No default test cycle. Makes point-edits without proactively running tests unless explicitly instructed — a workflow gap to close when wiring Spark into a goal-mode or otherwise autonomous setup (Help Net Security).
• Rollout visibility bugs. Access-tier and rate-limit reporting in the CLI and web UI showed inconsistencies during the research-preview period (see Agentic behavior notes); confirm current tier eligibility directly rather than trusting /status alone (GitHub issue #12992).

Related

• agent-harness — Spark's documented profile (fast, minimal-edit, developer-in-the-loop, no auto-verification, limited API reach) fits a fast-tier subagent role feeding a stronger reviewing model, not a lead role, in a harness built for long-horizon or multi-file work.
• GPT-5.5 — later-generation OpenAI coding model that, unlike Spark, runs its own test cycle by default and is generally API-addressable.
• SWE-bench Pro, Terminal-Bench — the two benchmarks Cerebras names qualitatively for Spark, with no numeric score published as of 2026-07-12.