Is GPT-5.4 better than Claude Opus 4.6?

It depends on the task. GPT-5.4 scores higher on computer use benchmarks (75% vs 72.7% on OSWorld) and costs half as much per token ($2.50/$15 vs $5/$25 per million tokens). Claude Opus 4.6 leads on agentic coding (65.4% on Terminal-Bench 2.0 vs GPT-5.4's 64.7%) and is widely considered better at sustained multi-file code work. For general professional workflows and cost efficiency, GPT-5.4 has the edge. For deep coding and agent orchestration, Opus still leads.

How much does GPT-5.4 cost compared to Claude Opus 4.6?

GPT-5.4 standard costs $2.50 per million input tokens and $15 per million output tokens. Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens. That makes GPT-5.4 roughly half the price per token. Both offer 1M token context windows with premium pricing above certain thresholds (272K for GPT-5.4, 200K for Opus 4.6). GPT-5.4 Pro costs $30/$180 per million tokens. Claude Opus 4.6 fast mode costs $30/$150 per million tokens.

Does GPT-5.4 have computer use like Claude?

Yes. GPT-5.4 is OpenAI's first general-purpose model with native computer use capabilities. It can operate a computer through both Playwright code and direct mouse/keyboard commands from screenshots. It scored 75% on OSWorld-Verified, surpassing the human baseline of 72.4%. Claude Opus 4.6 scored 72.7% on the same benchmark. Both models can now automate desktop workflows, but GPT-5.4 edges ahead on this specific benchmark.

Should I switch from Claude to GPT-5.4 for coding?

Not necessarily. Claude Opus 4.6 still leads on Terminal-Bench 2.0 (the most respected agentic coding benchmark) and SWE-bench Verified. Developers who work on complex multi-file projects generally still prefer Claude for sustained code quality and planning. GPT-5.4 is competitive and significantly cheaper, making it a strong choice for teams optimizing costs. Many developers are now using both - Claude for deep coding work, GPT-5.4 for general-purpose tasks and workflows where cost efficiency matters.

GPT-5.4 vs Claude Opus 4.6: Which Should Run Your Workflows?

GPT-5.4 dropped on March 5, 2026 - two days after GPT-5.3 Instant, in what's becoming OpenAI's strategy of keeping the news cycle permanently occupied. The headline pitch: the first general-purpose OpenAI model with native computer use, a 1M token context window, and a price point that undercuts Claude Opus by roughly half.

For anyone building AI into their workflows - whether that's coding, research, document processing, or agent automation - this is the first time in months that the "which model should I use?" question has genuinely interesting answers on both sides. Here's the honest comparison.

The benchmarks, without the spin

Both companies cherry-pick the benchmarks where they look best. Here's what the numbers actually say when you lay them side by side.

Computer use (OSWorld-Verified): GPT-5.4 hits 75.0%, surpassing the human baseline of 72.4%. Claude Opus 4.6 scores 72.7%. This is GPT-5.4's most impressive headline number - it's the first general-purpose model to beat humans on desktop navigation tasks involving screenshots and keyboard/mouse control. Opus is close behind but GPT-5.4 genuinely leads here.

Agentic coding (Terminal-Bench 2.0): Claude Opus 4.6 scores 65.4%. GPT-5.4 scores 64.7% - close, but Opus still holds the top spot on the benchmark that matters most to developers building complex software. SWE-bench Verified tells a similar story: Opus at 80.8% remains the model developers trust for sustained multi-file code work.

Professional knowledge work (GDPval): GPT-5.4 scores 83% across 44 occupations spanning the top 9 US GDP-contributing industries. This is OpenAI's own benchmark, so take it with appropriate skepticism, but the specific claims are notable - spreadsheet modeling jumped from 68.4% to 87.3%, and human raters preferred GPT-5.4's presentations 68% of the time over GPT-5.2's.

Hallucination rates: OpenAI reports GPT-5.4 is 33% less likely to make factual errors per claim and 18% less likely to produce responses containing any errors, compared to GPT-5.2. Note: the comparison is against 5.2, not 5.3 - a pattern worth catching. Anthropic hasn't published a direct hallucination comparison against GPT-5.4, but Claude has generally been the model that says "I'm not sure" instead of making something up with a straight face.

Context window: Both support 1M tokens. GPT-5.4's is available in the API immediately. Opus 4.6's is in beta for organizations in usage tier 4. In practice, both handle enormous codebases and document sets - the real-world difference is in how they handle retrieval and reasoning at the edges of that window, and neither company provides great data on degradation at scale.

Independent ranking: On the Artificial Analysis Intelligence Index, GPT-5.4 (xhigh reasoning) scores 57 and Opus 4.6 (max effort) scores 53. That puts GPT-5.4 slightly ahead on a composite of reasoning, knowledge, math, and coding - but both are well above the field.

The pricing gap is real

This is where GPT-5.4 makes the strongest case.

GPT-5.4 standard: $2.50 per million input tokens, $15.00 per million output tokens. Above 272K input tokens, input pricing doubles to $5.00.

Claude Opus 4.6: $5.00 per million input tokens, $25.00 per million output tokens. Above 200K input tokens (1M beta), pricing rises to $10.00 input and $37.50 output.

At standard rates, GPT-5.4 is half the price of Opus. That's not a marginal difference - for teams running thousands of API calls per day, it's the difference between a manageable line item and a budget conversation.

Both offer premium tiers: GPT-5.4 Pro at $30/$180 per million tokens, Claude Opus 4.6 fast mode at $30/$150. And both offer batch processing discounts (50% for both).

The cost comparison shifts when you factor in caching. Claude's prompt caching gives you 90% savings on cached input (reads at $0.50/MTok), which is more aggressive than GPT-5.4's automatic 50% cached input discount. For workflows with highly repetitive system prompts or context - exactly the kind of thing agent systems produce - Claude's caching advantage narrows the per-token gap significantly.

But if you're comparing sticker prices for standard API usage, GPT-5.4 is meaningfully cheaper.

What GPT-5.4 does that Claude doesn't

Native computer use in a general-purpose model. Claude has computer use capabilities, but they're accessed through the API with a containerized setup (Docker). GPT-5.4 bakes computer use directly into the model - it can write Playwright code to automate browsers and issue direct mouse/keyboard commands from screenshots, all within the same model context. For developers building agents that need to interact with desktop software, GPT-5.4's approach is simpler to integrate.

Tool Search. GPT-5.4 introduces a new API feature that lets the model receive a lightweight list of available tools and look up full definitions on demand, rather than stuffing all tool definitions into the system prompt. OpenAI reports this reduces prompt sizes by 47% on tool-heavy workflows. Claude doesn't have an equivalent - you still pass full tool definitions in every request.

ChatGPT for Excel. Launched alongside GPT-5.4, this puts AI directly inside spreadsheet workflows for financial modeling, analysis, and data manipulation. Claude has no native spreadsheet integration (though it works well with spreadsheet data via the API).

What Claude does that GPT-5.4 doesn't

Claude Code. Anthropic's command-line coding agent is widely considered the best coding assistant on the market. It reads your entire project, understands file relationships, modifies codebases, runs tests, and commits to GitHub. OpenAI's Codex is competitive but Claude Code's sustained performance on complex codebases - the kind of work where you need the model to hold context across dozens of files and make coordinated changes - still has a meaningful edge.

Cowork. Claude's desktop agent that can read your screen, navigate your files, and take actions on your computer in a persistent session. GPT-5.4's computer use is API-level; Claude's Cowork wraps it in a consumer-friendly desktop experience.

Projects and Skills. Claude's project system lets you give the model persistent context - documents, instructions, style guides - that it references across conversations. GPT-5.4 has custom GPTs and system prompts, but Claude's project architecture is deeper for teams that need consistent behavior across sessions.

The writing. This is subjective but widely shared: Claude produces noticeably better prose. Less corporate filler, fewer "let's dive in" patterns, more natural structure. GPT-5.4 improved its tone - OpenAI says they "loosened" it - but early reports confirm it still generates more generic-sounding output than Claude on writing-heavy tasks.

What developers are actually saying

The reaction from builders has been unusually split - which itself is the story.

Matt Shumer, who's been one of the most visible AI developers on X, called GPT-5.4 "the best model in the world, by far." His one caveat: frontend design still lags behind Opus 4.6 and Gemini 3.1 Pro. The Every team confirmed the benchmarks directionally: developers who were "90% Claude a month ago are now 50/50."

That 50/50 split is the real takeaway. For the first time since Claude Opus 4.5 launched, there's a genuine two-horse race for professional AI workflows. The old conventional wisdom - "Claude for coding, GPT for everything else" - is breaking down because GPT-5.4's coding is now competitive enough that the price difference starts to matter.

The early complaints are worth noting too. Users report GPT-5.4 occasionally "goes rogue on the details" - adding features nobody asked for (GDPR checkboxes on demo sites), leaking system prompts into UI elements, and just making stuff up with confidence. OpenAI explicitly said they "loosened" the model to be more conversational, which appears to have increased the rate of these creative misfires.

How to choose

Choose GPT-5.4 if: cost efficiency is a primary concern, you're building agents that need native computer use, your workflows are tool-heavy and would benefit from Tool Search, you do a lot of spreadsheet and presentation work, or you need the 1M context window without a beta waitlist.

Choose Claude Opus 4.6 if: you're doing sustained, complex coding across large codebases, you value writing quality and natural tone, your workflows benefit heavily from prompt caching (agent loops, repetitive context), you need Claude Code as your primary development environment, or you're already invested in Claude's project and skills ecosystem.

Use both if: you're serious about not creating a single point of failure with any AI provider - which, given recent events, is a more relevant consideration than it was a month ago. The models are close enough in capability that building an abstraction layer and routing different task types to different providers isn't just prudent - it's becoming the obvious architecture.

The bigger picture

Two months ago, the model hierarchy was clear: Claude for deep work, GPT for breadth, Gemini for multimodal. GPT-5.4 scrambles that hierarchy by being genuinely competitive on Claude's home turf (coding and professional reasoning) while maintaining OpenAI's traditional strengths in tool use and ecosystem breadth.

What that means if you're building AI workflows: the cost of switching models is dropping, the capability gaps are narrowing, and the real differentiators are increasingly about ecosystem (Claude Code vs Codex, Cowork vs computer use API) rather than raw intelligence.

The best model for your workflow in March 2026 might not be the best model for your workflow in June. Build accordingly.

This is part of a series covering AI tools and the ecosystem around them. See also: What Anthropic's Supply Chain Risk Label Means If You Build on Claude, Perplexity Computer vs Claude Cowork, How Much Does Perplexity Computer Cost?, and Best OpenClaw Alternatives That Don't Require Coding.