The “One Model” Pitch OpenAI Has Been Building Toward
For the past 18 months, using AI for serious work meant juggling models. One for code. Another for research. A third for anything that required operating a computer. OpenAI’s answer, shipped on March 5, 2026: merge everything into a single system and call it GPT-5.4.
The pitch is simple. GPT-5.4 absorbs the coding engine from GPT-5.3-Codex, adds native computer-use capabilities (a first for any OpenAI general-purpose model), supports a 1-million-token context window in the API, and ships simultaneously across ChatGPT, the API, and Codex. No model-switching. No separate agents.
The real question is whether the consolidation holds up under scrutiny, or whether “good at everything” translates to “best at nothing.”
Knowledge Work: The 83% GDPval Number, Unpacked
OpenAI’s headline stat comes from GDPval, a benchmark spanning nine major U.S. industries and 44 occupations. Tasks mirror daily professional work: building financial models, scheduling hospital shifts, assembling sales decks. Outputs are blind-evaluated by real practitioners against human-produced work.
GPT-5.4 hit 83.0% on GDPval, meaning industry professionals judged the AI output as matching or exceeding human peer quality in roughly eight out of ten comparisons. GPT-5.2 sat at 70.9%, a gap of about 13 percentage points (source: OpenAI’s official GPT-5.4 announcement, March 2026).
The jump is most visible in spreadsheet modeling. Simulating junior investment banking analyst tasks, GPT-5.4 averaged 87.3% versus GPT-5.2’s 68.4%, a roughly 19-point improvement (source: OpenAI GDPval results). On the legal side, Harvey’s BigLaw Bench returned a 91% score. Mercor’s APEX-Agents benchmark, which tests professional skills in law and finance, also ranked GPT-5.4 first, according to a statement from Mercor CEO Brendan Foody reported by TechCrunch (March 6, 2026).
On hallucination reduction: OpenAI reports GPT-5.4’s individual factual claims are 33% less likely to be false compared to GPT-5.2, and full responses are 18% less likely to contain any errors (source: OpenAI announcement).
These numbers are self-reported by OpenAI. Independent replication from third-party labs is still pending. That said, the direction and magnitude are consistent across multiple evaluation frameworks.
Coding: Consolidation Without Compromise (Mostly)
The big structural change: GPT-5.3-Codex’s coding capabilities are now baked into the mainline model. Developers no longer need a dedicated coding model.
On SWE-Bench Pro, which tests real-world software engineering tasks, GPT-5.4 scored 57.7%. GPT-5.3-Codex got 56.8%. GPT-5.2 trailed at 55.6% (source: OpenAI benchmark results). The merge didn’t cause regression. It marginally improved scores while bundling general-purpose reasoning and computer-use capabilities alongside the code engine.
AI evaluation writer Dan Shipper noted after early access that GPT-5.4 showed notably strong planning ability for long tasks and solid code review, while costing roughly half of what Claude Opus charges per API call. Wharton professor Ethan Mollick used a single prompt to generate a 3D scene inspired by the novel Piranesi, comparing the output to GPT-4’s attempt from two years ago. The quality gap was immediately apparent.
A new experimental feature, Playwright Interactive, lets GPT-5.4 visually debug web and Electron applications in real time. The model writes code, opens a browser, tests the output, and iterates. In OpenAI’s demo, a single lightweight prompt produced a complete isometric theme park simulation game with tile-based pathways, visitor AI pathfinding, queuing behavior, and four real-time metrics. The model handled both development and QA.
In Codex, the `/fast` mode pushes token generation speed up to 1.5x, which matters for iterative coding sessions where latency kills flow.
Where Coding Gets Complicated
Here’s the nuance the headline benchmarks obscure. On SWE-Bench Verified (the broader, more commonly cited coding eval), Claude Opus 4.6 leads with 80.8%, followed closely by Gemini 3.1 Pro at 80.6%. GPT-5.4 has not published a directly comparable SWE-Bench Verified score in OpenAI’s announcement materials, focusing instead on SWE-Bench Pro. The distinction matters. SWE-Bench Verified is the most widely referenced coding benchmark in the industry, and Opus currently owns the top spot on it (source: Anthropic’s Opus 4.6 benchmarks and DigitalApplied comparison, March 2026).
For pure competitive coding, Gemini 3.1 Pro holds the highest LiveCodeBench Pro Elo at 2887 (source: DigitalApplied’s three-way comparison, March 2026).
So: GPT-5.4’s coding is strong, arguably the best all-in-one package. But if raw coding benchmarks are your only metric, Opus and Gemini still lead on specific leaderboards.
Computer Use: The Feature That Actually Changes Things
This is the headline capability. GPT-5.4 is the first OpenAI general-purpose model with native computer-use abilities built directly into the reasoning engine, not bolted on as a separate module.
On OSWorld-Verified, which tests desktop navigation through screenshots and mouse/keyboard interactions, GPT-5.4 scored 75.0%. The human baseline is 72.4%. GPT-5.2 managed 47.3% (source: OpenAI announcement, citing the OSWorld paper for the human baseline).
That is not a typo. The model outperforms the average human at navigating a computer desktop using screenshots alone.
On Online-Mind2Web, a browser-only navigation benchmark using screenshots, GPT-5.4 hit 92.8%, compared to 70.9% for ChatGPT Atlas’s Agent Mode (source: OpenAI benchmark results).
A real-world deployment validates this: Mainstay used GPT-5.4 for automated form filling across roughly 30,000 property tax portals. First-attempt success rate hit 95%, with 100% success within three attempts. The prior best from comparable models was 73% to 79%. Session completion speed tripled, and token consumption dropped about 70% (source: OpenAI case study in the GPT-5.4 launch materials).
The Vision Upgrade Behind It
Computer use depends on seeing interfaces accurately. GPT-5.4 introduces an “original” image input mode supporting up to 10.24 megapixels or 6,000-pixel maximum edge length. This is a meaningful upgrade for tasks where small UI elements, dense spreadsheets, or fine-print text need to be read precisely.
The 1-Million-Token Context Window: Useful, But Read the Fine Print
GPT-5.4’s API and Codex versions support up to 1.05 million tokens (922K input, 128K output). For developers working with large codebases or legal document sets, this is a genuine capability expansion.
But context length does not equal context quality. OpenAI’s own long-context evaluations show performance degradation at extreme lengths. On the MRCR needle-retrieval benchmark, GPT-5.4 holds 86.0% accuracy through 128K tokens but drops to 36.6% at the 512K to 1M range (source: OpenAI technical documentation). On Graphwalks BFS, accuracy drops from 93.0% at the 0-128K range to 21.4% at 256K-1M (source: OpenAI benchmarks).
Translation: the 1M context window exists, and it works for some retrieval tasks. It is not a magic “load everything and get perfect answers” solution. Treat the 128K range as the reliable sweet spot.
Also worth noting: prompts exceeding 272K input tokens are billed at 2x the standard input rate and 1.5x output rate for the entire session (source: OpenAI pricing documentation). Long context gets expensive fast.
Pricing: The $80 “Hi” Meme and What It Actually Costs
The viral reaction on Chinese social media was that saying “Hi” to GPT-5.4 burns $80. That is an exaggeration, but not entirely baseless if you are thinking about GPT-5.4 Pro on extended reasoning tasks with massive context.
Here is the actual pricing structure:
API pricing (per 1M tokens):
- GPT-5.4 Standard: $2.50 input / $15.00 output
- GPT-5.4 Cached input: $0.625
- GPT-5.4 Pro: $30.00 input / $180.00 output
- Batch and Flex pricing: 50% of standard rate
ChatGPT subscription access:
- Plus ($20/month): GPT-5.4 Thinking, 80 messages per 3 hours
- Pro ($200/month): Unlimited GPT-5.4 Pro access with dedicated GPU slice
For context, Artificial Analysis reported it cost $2,950.85 to evaluate GPT-5.4 (xhigh) on their full Intelligence Index, during which the model generated 120 million output tokens (source: Artificial Analysis, March 2026). The model is verbose. Budget accordingly.
GPT-5.2 Thinking remains available in Legacy Models until June 5, 2026, then gets retired permanently.
GPT-5.4 vs Claude Opus 4.6 vs Gemini 3.1 Pro: The Three-Way Comparison
| Benchmark / Metric | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|
| GDPval (Knowledge Work) | 83.0% | 78.0% | N/A |
| OSWorld-Verified (Desktop Use) | 75.0% | 72.7% | N/A |
| SWE-Bench Verified (Coding) | 57.7% (Pro variant) | 80.8% | 80.6% |
| ARC-AGI-2 (Abstract Reasoning) | 73.3% | 75.2% | 77.1% |
| GPQA Diamond (Science) | 92.8% | 91.3% | 94.3% |
| BrowseComp (Web Research) | 82.7% | 84.0% | N/A |
| Context Window (API) | 1.05M tokens | 200K (1M beta) | 2M tokens |
| API Input Price (per 1M tokens) | $2.50 | $5.00 | $2.00 |
| API Output Price (per 1M tokens) | $15.00 | $25.00 | $12.00 |
| Artificial Analysis Intelligence Index | 57 | 53 | 57 |
Sources: OpenAI GPT-5.4 announcement (March 2026), Anthropic Opus 4.6 benchmarks (February 2026), Google DeepMind Gemini 3.1 Pro release (February 2026), Artificial Analysis Intelligence Index (March 2026), DigitalApplied three-way comparison (March 2026).
Reading the table:
- GPT-5.4 owns knowledge work automation and computer use. If your workflow involves operating software, filling forms, navigating interfaces, or producing professional documents, it is the strongest option available today.
- Claude Opus 4.6 owns production-grade coding and code quality. Developers doing serious SWE work, refactoring, and bug-fixing will find Opus more reliable on the benchmarks that matter most.
- Gemini 3.1 Pro owns abstract reasoning and price-performance. At $2/$12 per million tokens with a 2M context window and 94.3% on GPQA Diamond, it is the most cost-effective frontier model for research-heavy and reasoning-heavy workloads.
The Artificial Analysis Intelligence Index, a composite score across reasoning, knowledge, math, and coding, currently ties GPT-5.4 and Gemini 3.1 Pro at 57, with Claude Opus 4.6 at 53 (source: Artificial Analysis, March 2026). Benchmark convergence at the frontier is arguably the real story of 2026.
Tool Search: The API Feature Developers Should Watch
A detail buried in the launch materials deserves attention. GPT-5.4 introduces Tool Search, a mechanism that lets the model receive a lightweight list of available tools and look up full definitions on demand, instead of loading every tool definition into every prompt.
In testing on 250 tasks from Scale’s MCP Atlas benchmark with 36 MCP servers enabled, Tool Search reduced total token usage by 47% while maintaining accuracy (source: OpenAI announcement, confirmed by Tom’s Guide reporting, March 2026).
For developers building agentic systems with many integrations, this is a direct cost savings mechanism. It also reduces latency, since smaller prompts mean faster initial processing.
The Honest Verdict
GPT-5.4 is not the best at any single thing. It is the best at doing many things inside one model without forcing developers to route between specialized systems.
Choose GPT-5.4 if: Your workflow spans coding, document production, computer operation, and research within the same pipeline. The consolidation value is real. It eliminates model-switching overhead and simplifies agent architectures.
Choose Claude Opus 4.6 if: You are a developer whose primary concern is code quality, SWE task completion, and deep reasoning on complex codebases. Opus remains the SWE-Bench leader and produces higher-quality code outputs in expert evaluations.
Choose Gemini 3.1 Pro if: You need the strongest abstract reasoning at the lowest cost, or you work with very long documents and need production-grade 2M context. At $2/$12 per million tokens, it is 2.5x cheaper than GPT-5.4 on output and 2x cheaper than Opus.
The power move for teams with engineering capacity: Route tasks to the model best suited for each step. Use GPT-5.4 for computer use and professional document tasks, Opus for complex code, and Gemini for bulk reasoning and long-context retrieval. The frontier models of March 2026 are close enough in overall capability that the routing strategy matters more than the model choice.
GPT-5.4 is not a knockout punch. It is a consolidation play, and a strong one. Whether the premium pricing holds up against Gemini’s aggressive cost structure and Opus’s coding lead will depend on how much developers value the “one model for everything” convenience versus best-in-class performance on individual tasks.
