Bottom line
Route by bottleneck, not by benchmark headline.
Spark is the fast interactive Codex model. GPT-5.5 is the stronger OpenAI model to reach for when correctness, repo-wide reasoning, or autonomous execution matters. Claude Sonnet 4.6 is Anthropic's current Sonnet tier and a strong daily-driver choice when Claude Code, Anthropic APIs, 1M context, or adaptive thinking fit the job.
Snapshot
The useful stats
| Axis | GPT-5.3-Codex-Spark | GPT-5.5 | Claude Sonnet 4.6 |
|---|---|---|---|
| Best role | Real-time Codex collaboration | Frontier Codex and professional work | Fast, capable Claude agent backbone |
| Public status | Research preview | Public API model page plus ChatGPT and Codex rollout | Current Sonnet model |
| Context | 128k | 400k in Codex; 1,050,000 in API | 1M |
| Inputs | Text-only at launch | Text and image input; text output | Text and image input |
| Speed signal | More than 1000 tokens/sec on ultra-low-latency Cerebras serving | Fast mode in Codex is 1.5x faster for 2.5x cost | Fast comparative latency in Claude docs |
| Pricing signal | No general API price found; separate preview limits | $5/M input, $0.50/M cached input, $30/M output; long prompts over 272k input tokens cost more | $3/M input, $15/M output |
| Default behavior | Lightweight, targeted; does not auto-run tests unless asked | API reasoning effort supports none, low, medium default, high, and xhigh | Fine-grained thinking effort, extended thinking, and adaptive thinking |
Decision ladder
When to reach for each model
| Situation | First reach | Why |
|---|---|---|
| You are watching the agent and want fast iteration | GPT-5.3-Codex-Spark | The model is designed for real-time coding, interrupts, redirects, and rapid steering. |
| You need one small patch in a familiar file | GPT-5.3-Codex-Spark | Its lightweight default is a feature when the desired diff is narrow. |
| You need a fast review or scout pass before asking a bigger model | GPT-5.3-Codex-Spark | Low latency makes it useful for second opinions and quick triage. |
| You need the agent to inspect a large repo, reason through tradeoffs, run tests, and land a change | GPT-5.5 | The task is quality-bound and tool-bound rather than latency-bound. |
| You need a hard architecture call, difficult bug, or high-confidence review | GPT-5.5 | OpenAI positions it as the stronger professional and Codex tier. |
| You are in Claude Code or building against Anthropic APIs | Claude Sonnet 4.6 | It is Anthropic's current speed/intelligence workhorse with 1M context. |
| You want the strongest Anthropic model for deep refactors or multi-agent coordination | Claude Opus, not Sonnet | Anthropic says Opus remains the deepest-reasoning choice for those jobs. |
Benchmarks
Stats are directional, not a clean scoreboard.
| Metric family | GPT-5.3-Codex-Spark | GPT-5.5 | Claude Sonnet 4.6 |
|---|---|---|---|
| SWE-style coding | OpenAI says strong SWE-Bench Pro performance in a fraction of GPT-5.3-Codex time, but the public page did not expose a numeric score in text. | 58.6% on SWE-Bench Pro Public in OpenAI's table. | Anthropic system card reports 79.6% SWE-bench Verified; this is not the same benchmark. |
| Terminal / agentic coding | OpenAI says strong Terminal-Bench 2.0 performance faster than GPT-5.3-Codex, no numeric text found. | 82.7% on Terminal-Bench 2.0. | Anthropic reported scores vary by harness and thinking setup; compare carefully. |
| Computer use | Not the right target; Spark is text-only at launch. | 78.7% OSWorld-Verified in OpenAI's table. | Anthropic positions Sonnet 4.6 as much improved for computer use. |
| Tool use | Useful in Codex's real-time loop, but public source emphasizes latency more than tool benchmark scores. | 75.3% MCP Atlas and 55.6% Toolathlon in OpenAI's table. | Strong tool-selection and error-correction positioning; use local evals for direct comparison. |
Playbook
A simple routing policy
- Start with Spark when the task should fit in one tight interactive loop.
- Escalate to GPT-5.5 when Spark starts asking for broad context, proposes a risky diff, or needs to own verification.
- Use Sonnet 4.6 when Claude Code is already the best product surface, when 1M-context Sonnet economics matter, or when local evals show stronger instruction following on your codebase.
- Route critical refactors and production-risk reviews to the most capable model available, then use a faster model for follow-up cleanup after the hard part is settled.
- Keep local scorecards. Vendor stats choose the first model; your repo's failure modes decide the durable default.
Caveats
What can go stale
- Spark access and limits are research-preview facts.
- GPT-5.5 now has a public API model page, but long-context pricing and tool-call fees can materially change automation cost.
- Sonnet means Sonnet 4.6 in this guide because that was the current Anthropic Sonnet page and model table on 2026-05-09.
- Do not compare Sonnet 4.6 SWE-bench Verified directly to GPT-5.5 SWE-Bench Pro Public. They are different benchmarks.
- Public pages do not capture every product-plan limit, enterprise admin setting, or Codex/Claude Code UI behavior.
Receipts
Claim table
| Where | Claim | Confidence | Source | Caveat |
|---|---|---|---|---|
| Bottom line | GPT-5.3-Codex-Spark is a research-preview, smaller Codex model designed for real-time coding in Codex. | High | OpenAI | Research-preview access and limits may change quickly. |
| Snapshot | Spark is text-only with a 128k context window at launch. | High | OpenAI | OpenAI says later ultra-fast models may add larger context and multimodal input. |
| Stats | Spark is optimized for latency and can deliver more than 1000 tokens per second on ultra-low-latency Cerebras serving. | High | OpenAI | The speed claim is hardware and serving-path dependent. |
| Snapshot | Spark keeps a lightweight default working style and does not automatically run tests unless asked. | Medium | OpenAI | This is release-note behavior, not a guaranteed invariant across all Codex surfaces. |
| Snapshot | GPT-5.5 is available in ChatGPT and Codex, with Codex availability across Plus, Pro, Business, Enterprise, Edu, and Go plans and a 400k Codex context window. | High | OpenAI | Subscription and admin settings may affect access. |
| Stats | GPT-5.5 is listed on the official API model page at $5/M input, $0.50/M cached input, and $30/M output, with a 1,050,000-token context window, 128,000 max output tokens, text and image input, text output, and reasoning effort values of none, low, medium default, high, and xhigh. | High | OpenAI API docs | Long prompts over 272k input tokens have higher pricing; check the current pricing page before large-context automation. |
| Benchmarks | GPT-5.5 scores 58.6% on SWE-Bench Pro Public and 82.7% on Terminal-Bench 2.0 in OpenAI's published table. | Medium | OpenAI | Vendor benchmark tables are not harness-normalized against Spark or Sonnet. |
| Benchmarks | GPT-5.5 scores 78.7% on OSWorld-Verified and 75.3% on MCP Atlas in OpenAI's published table. | Medium | OpenAI | Computer-use and tool-use results depend heavily on scaffold and allowed tools. |
| Snapshot | Current Claude Sonnet means Claude Sonnet 4.6 as of 2026-05-09. | High | Anthropic API docs | Anthropic may ship a newer Sonnet later; re-check before relying on this guide. |
| Snapshot | Claude Sonnet 4.6 is positioned as the best combination of speed and intelligence, with 1M context, 64k max output, fast comparative latency, extended thinking, adaptive thinking, and $3/M input plus $15/M output pricing. | High | Anthropic API docs | Third-party platform routing and pricing may vary. |
| Decision ladder | Anthropic positions Sonnet 4.6 for coding, agents, computer use, tool selection, error correction, and production workflows. | High | Anthropic | Positioning is vendor-supplied. |
| Decision ladder | Anthropic still recommends Opus for tasks demanding the deepest reasoning, including codebase refactoring and multi-agent coordination where precision is paramount. | High | Anthropic | This comparison is within Anthropic's own model family, not against GPT-5.5. |
| Benchmarks | Direct GPT-5.3-Codex-Spark vs GPT-5.5 vs Sonnet benchmark comparisons should be treated cautiously unless harness, effort setting, tools, sample count, prompt, and budget are normalized. | High | OpenAI | This is an inference from vendor benchmark heterogeneity, not a single published benchmark result. |
Sources
How this was built
| Publisher | Source | Type | Accessed |
|---|---|---|---|
| OpenAI | Introducing GPT-5.3-Codex-Spark | Official | 2026-05-09 |
| OpenAI | Introducing GPT-5.5 | Official | 2026-05-09 |
| OpenAI API docs | GPT-5.5 model page | Official | 2026-05-09 |
| Anthropic | Claude Sonnet 4.6 | Official | 2026-05-09 |
| Anthropic | Introducing Claude Sonnet 4.6 | Official | 2026-05-09 |
| Anthropic API docs | Models overview | Official | 2026-05-09 |
| Anthropic | Claude Sonnet 4.6 System Card | Official | 2026-05-09 |