Your personal memory, across sessions, agents, and devices.

Codex Spark: What It Was and What Replaced It

memU Team memU Team

GPT-5.3-Codex-Spark was OpenAI's first model built for real-time coding. It launched on February 12, 2026 as a research preview for ChatGPT Pro users. It ran on Cerebras hardware at more than 1,000 tokens per second. OpenAI retired it on September 14, 2026. If you still have it pinned in a Codex config, that config now points at a model that does not exist. Below is how Spark got its speed and what it could do. Fast mode and Ultrafast mode now fill the gap it left, and the last sections cover both.

What was Codex Spark, in one sentence?

Codex Spark (GPT-5.3-Codex-Spark) was a smaller version of GPT-5.3-Codex that OpenAI served on Cerebras' Wafer Scale Engine 3 so Codex could answer in near real time.

OpenAI's launch post called it "our first model designed for real-time coding." It was text-only with a 128k context window. Cerebras' own announcement described it as the first release in the Cerebras and OpenAI collaboration. OpenAI positioned it as a second model for quick, interactive edits, used alongside GPT-5.3-Codex.

How did Spark differ from the full Codex model and today's speed modes?

Spark was a separate, smaller model. Today's speed options in Codex run the regular models faster in exchange for more usage.

GPT-5.3-Codex-Spark Fast mode Ultrafast mode
What it is A smaller model on Cerebras hardware A faster service tier for supported models The fastest service tier for GPT-6 Astra and GPT-6.1 Sol
Speed claim (OpenAI) More than 1,000 tokens per second 1.5x for GPT-5.6 Up to 8x faster for GPT-6 Astra in Codex
Who could use it ChatGPT Pro, research preview ChatGPT sign-in in the desktop app, CLI and IDE extension Pro $500 and eligible Enterprise and Edu plans
Cost in Codex Its own rate limit, separate from standard limits 2.5x the Standard rate against included limits 8x the Standard rate against included limits
Status Retired September 14, 2026 Available Available. GPT-6.1 Sol Ultrafast came to Codex on October 8, 2026

Sources: OpenAI's launch post and Codex changelog, plus the Codex speed docs.

How did Codex Spark work?

Spark's speed came first from hardware. OpenAI added a low-latency serving path on Cerebras chips to the same production stack as the rest of its fleet. The launch post says GPUs stay "foundational" for training and inference, with Cerebras covering work where latency matters most.

OpenAI also found that token speed was only part of the delay. It added a persistent WebSocket connection and tuned the Responses API. OpenAI says this cut the overhead per roundtrip by 80% and per-token overhead by 30%. Time to first token fell by 50%. The WebSocket path was on by default for Spark, and OpenAI said it would become the default for all models. The WebSocket path outlived Spark: OpenAI's Ultrafast API guide now strongly recommends it for agents that make many tool calls in a row.

Spark also behaved differently by design. OpenAI said it made small targeted edits and skipped tests unless asked. You could interrupt and redirect it mid-task. That fits changing a layout or renaming a function better than a multi-hour refactor.

How capable was Spark compared with full Codex?

OpenAI said Spark showed "strong performance" on SWE-Bench Pro and Terminal-Bench 2.0. It finished tasks in a fraction of the time GPT-5.3-Codex needed. Cerebras' post put it above GPT-5.1-Codex-mini on those benchmarks. Both are vendor claims. OpenAI published the results as charts with no independent reproduction, so we do not quote exact scores here.

Spark traded depth for speed. Its small window and lighter defaults made it a fit for short loops where you watch every change. For long autonomous work, OpenAI's own launch post credits its larger frontier models, which it says can work for hours or days.

Who could use Spark, and what did it cost?

Spark rolled out to ChatGPT Pro users in the Codex app, the Codex CLI and the VS Code extension. In the CLI you selected it with codex --model gpt-5.3-codex-spark or the /model command. The changelog notes it was not in the public API at launch. A small group of design partners got API access.

Usage ran on its own rate limit and did not count against standard Codex limits. OpenAI warned that access could queue or slow down when demand was high, because the specialized hardware had limited capacity.

Why was Codex Spark retired, and what should you use now?

OpenAI deprecated GPT-5.3-Codex-Spark on September 14, 2026, according to the Codex changelog. The research preview is gone from every Codex client, and OpenAI tells users to update any saved configuration or custom agent that still selects it. It points to Fast mode instead and gives no reason for the retirement.

Speed in Codex now comes from service tiers on the main models. The speed docs describe two:

  • Fast mode works with GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna where available. Toggle it with /fast in the CLI. It uses included limits at 2.5x the Standard rate.
  • Ultrafast mode supports GPT-6 Astra and GPT-6.1 Sol. OpenAI says GPT-6 Astra Ultrafast generates tokens up to 8x faster than Standard in Codex. It uses included limits at 8x the Standard rate.

The models page lists Spark under deprecated Codex models next to GPT-5.5, which left Codex for ChatGPT sign-in on October 14, 2026. If you run Codex with an API key, availability follows your API organization instead.

What should you read in the fine print?

Every speed figure in this post comes from OpenAI or Cerebras. The 1,000 tokens per second figure is OpenAI's own. So are the latency cuts and the 8x Ultrafast claim. We found no independent benchmark of Spark inside Codex.

Spark lasted seven months, from February 12 to September 14, 2026. In that time OpenAI also retired GPT-5.4 and GPT-5.5 from Codex for ChatGPT sign-in. Anything in your setup that names a specific model needs a review every few months.

The bottom line

Codex Spark is gone. Remove gpt-5.3-codex-spark from config.toml and from any custom agent or script. For quick interactive edits, turn on Fast mode with a current model and watch the 2.5x usage rate. Ultrafast is worth testing only if you are on a plan that includes it and latency is your real bottleneck.

How do you keep coding agent memory when the model is retired?

When a model is retired, your agent's project knowledge has to live somewhere the retirement cannot reach. Codex keeps that knowledge in two places. AGENTS.md holds rules you write yourself. The optional Codex memories feature writes generated memory files to ~/.codex/memories/ on each machine. OpenAI calls those files generated state and advises against hand-editing them as your main control. Those files stay on one machine and inside Codex, so Claude Code or Cursor in the same repository starts without them.

memU keeps that layer outside any one model or host. A scheduled background task slices new session logs into jobs. The agent then distills them into memory and skill Markdown that you can read and edit. memU makes no LLM calls itself. It stores and embeds what the agent wrote, then retrieves it on request. A standing instruction in AGENTS.md or CLAUDE.md tells the agent to retrieve before answering. Codex and Claude Code share one backend, so lessons from one carry over to the other. Our post on why memU leaves the thinking to the agent explains the design.

The hosted version is free and syncs across devices. Get an API key at memu.so and send your agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

Self-hosted setup and the host support matrix are in the memU GitHub repository.

FAQ

Is Codex Spark still available?

No. OpenAI deprecated GPT-5.3-Codex-Spark on September 14, 2026. It is gone from the ChatGPT desktop app and from the Codex CLI and IDE extension.

How fast was GPT-5.3-Codex-Spark?

OpenAI said Spark delivered more than 1,000 tokens per second on Cerebras' Wafer Scale Engine 3. That is a vendor figure from the February 12, 2026 launch post.

What replaced Codex Spark?

OpenAI points users to Fast mode on current models such as GPT-6.1 Sol. Ultrafast mode is faster still. It covers Pro $500 plus eligible Enterprise and Edu plans, and GPT-6.1 Sol Ultrafast arrived in Codex on October 8, 2026.

What context window did Codex Spark have?

Spark launched with a 128k context window and accepted text only. OpenAI promised larger models and longer context for the family. Multimodal input was also on its list.

Was Codex Spark available in the API?

Not publicly. At launch OpenAI gave API access only to a small set of design partners and told API-key users to keep using gpt-5.2-codex.