Compare Claude Opus 4.7 vs GPT-5.4 vs Gemini 3.1 Pro on the Artificial Analysis Intelligence Index and you get a three-way tie at exactly 57 points. If that single number settled anything, you could stop reading here.
Released within 60 days of each other — Gemini 3.1 Pro on February 19, GPT-5.4 on March 5, and Claude Opus 4.7 on April 16 — all three sit at the top of the same composite leaderboard. The tie hides radically different workload winners, 3x speed gaps, and a pricing story Anthropic buried three paragraphs deep on its own docs page.
This comparison cuts past the benchmark reprints to the question that actually matters: which model should your team route which workload to, and what will it cost when the bill arrives.
Where each model wins (and loses)
The Intelligence Index tie is real — but it flattens sharp workload-specific divergence. Opus 4.7 dominates agentic coding and tool orchestration. GPT-5.4 Pro owns web research. Gemini 3.1 Pro leads on speed and price. On graduate-level science, all three hit a ceiling so close it doesn’t matter which you pick.
| Benchmark | Opus 4.7 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|
| SWE-bench Verified | 87.6% | ~78-80% | 80.6% |
| SWE-bench Pro | 64.3% | 57.7% | 54.2% |
| MCP-Atlas (tool orchestration) | 77.3% | 67.2% | 69.2% |
| BrowseComp (web research) | 79.3% | 89.3% (Pro) | 85.9% |
| GPQA Diamond | 94.2% | 92.8% | 94.3% |
| Humanity’s Last Exam (tools) | 54.7% | 52.1% | 51.4% |
| OSWorld-Verified | 78.0% | 75.0% | — |
Opus 4.7 leads SWE-bench Pro by 6.6 points over GPT-5.4 and 10.1 over Gemini — a decisive win on multi-language coding. MCP-Atlas is an 8-to-10-point lead on multi-server tool coordination. Per Vellum AI’s benchmark analysis, this is the widest gap Anthropic has held on any coding benchmark in a year.
The flip side is BrowseComp. GPT-5.4 Pro’s 89.3% sets a new state of the art — a full 10 points above Opus 4.7. That’s Opus’s most prominent public loss, and it’s not close. As The AI Corner summarized it: “Opus 4.7 extends the gap rather than ceding ground in agentic and coding workloads, though it underperforms on Terminal-Bench 2.0 and BrowseComp compared to competitors.” For a deeper look at how tool choice compounds these gaps, see our coding assistant comparison.
The real price of Claude Opus 4.7 vs GPT-5.4 vs Gemini 3.1 Pro
Before the pricing table: read this sentence from Anthropic’s pricing docs, three paragraphs into the page. “Opus 4.7 uses a new tokenizer compared to previous models… This new tokenizer may use up to 35% more tokens for the same fixed text.”
Translation: per-token pricing is unchanged at $5/$25. Per-task pricing quietly rose 0–35%, with code and structured data hit hardest. Most comparison write-ups have reprinted “$5/$25 unchanged” and stopped there. It isn’t unchanged in any way that matters to your bill.
| Model | Input $/MTok | Output $/MTok | Context | Effective Input* |
|---|---|---|---|---|
| Opus 4.7 | $5.00 | $25.00 | 1M | $5.00–$6.75 |
| Sonnet 4.6 | $3.00 | $15.00 | 1M | $3.00 |
| Haiku 4.5 | $1.00 | $5.00 | 200K | $1.00 |
| GPT-5.4 Standard | $2.50 | $15.00 | 272K | $2.50 |
| GPT-5.4 Pro | $30.00 | $180.00 | 272K | $30.00 |
| Gemini 3.1 Pro (<200K) | $2.00 | $12.00 | 1M | $2.00 |
| Gemini 3.1 Pro (>200K) | $4.00 | $18.00 | 1M | $4.00 |
*Effective input cost accounts for Opus 4.7’s tokenizer inflation (1.0–1.35x on identical text). Gemini pricing per Google’s Gemini API pricing page.
Four workloads make the gap concrete:
- Daily coding agent (1M input / 200K output per day): roughly $405/month on Opus 4.7 after tokenizer inflation — versus $180 on Sonnet 4.6, $165 on GPT-5.4 Standard, and $132 on Gemini 3.1 Pro. Finout’s analysis pegged an agent that ran $300/month on Opus 4.6 at $405/month on 4.7 with zero workload change.
- RAG assistant (5M input / 500K output monthly, 70% cache hits): about $652 on Opus 4.7 versus $392 on Sonnet 4.6 — a 40% savings our Sonnet 4.6 review covers in detail.
- Long-context work above 200K tokens: Claude wins on math. Gemini doubles to $4/$18 at the tier break, GPT-5.4 jumps to $5/MTok input above 272K, and Claude keeps flat pricing across the full 1M window.
- Deep web research session (100K output tokens): GPT-5.4 Pro costs roughly $18 per run versus $2.50 for Opus. The question is whether 10 extra BrowseComp points justifies a 7x multiplier.
The practical recommendation: measure actual token consumption on your own traffic before upgrading from 4.6 to 4.7. Headline per-token rates stopped being apples-to-apples across Claude versions on April 16.
Speed and verbosity: the hidden Opus tax
Per Artificial Analysis’s independent speed testing: Gemini 3.1 Pro runs at 132.3 tokens/second, GPT-5.4 xhigh at 79.5 t/s, and Opus 4.7 at 45.5 t/s. Opus ranks #76 of 133 models on throughput — below the 62.9 t/s median. Gemini finishes the same output in roughly a third the wall-clock time.
Verbosity compounds the slowness. During the Intelligence Index eval, Opus 4.7 generated 102 million output tokens versus a 35 million average across models. The model says more to accomplish the same thing, which means the speed penalty and the output-token bill grow together.
[Claude Opus 4.7 is] amongst the leading models in intelligence, but particularly expensive when comparing to other models of similar price. It’s also slower than average and very verbose.
Artificial Analysis
Caveat for Gemini: its time-to-first-token is 30.28 seconds — the highest in the peer group. Gemini front-loads latency, which matters for interactive chat. But once it starts talking, it finishes first. For any production endpoint where a human is waiting on the response, Opus’s benchmark wins don’t translate to user-experience wins.
Variants that actually matter
Most teams reach for “Opus” or “GPT-5.4” by reflex. The right answer is usually a tier below — which is where the real cost savings and throughput live.
Anthropic
- Opus 4.7 ($5/$25): genuine agentic complexity, long-running multi-step workflows, hardest coding tasks.
- Sonnet 4.6 ($3/$15): Anthropic itself positions this as handling 90%+ of coding tasks without compromise.
- Haiku 4.5 ($1/$5, 97.2 t/s, 73.3% SWE-bench Verified): high-volume and latency-sensitive work — matches Sonnet 4’s mid-2025 coding performance at a fifth the price.
OpenAI
- GPT-5.4 Standard ($2.50/$15): default for most use cases, per OpenAI’s GPT-5.4 announcement.
- GPT-5.4 Pro ($30/$180, Responses API only): reserve for web research where BrowseComp accuracy justifies the 12x input premium.
- GPT-5.4 Mini ($0.75/$4.50): throughput and high-volume coding via Codex.
- GPT-5.4 Thinking: ChatGPT-only, no separate API price. GPT-Rosalind (April 17, 2026) is a specialized life-sciences model with restricted access (Amgen, Moderna, Allen Institute, Thermo Fisher) — not a GPT-5.4 variant. “GPT-5.4-Cyber” does not appear in any OpenAI documentation.
- Gemini 3.1 Pro ($2/$12, doubles above 200K): long-context and price-sensitive flagship work. Cheaper than your daily latte per million tokens — just watch the 200K cliff.
- Gemini 3.1 Flash-Lite ($0.25/$1.50): highest-volume tasks where flagship quality is overkill.
Use-case routing: which model for which workload
Route by workload, not by vendor loyalty. See choosing your agent stack for the framework-level version of this decision.
| Workload | Pick | Why |
|---|---|---|
| Agentic coding (multi-step, tool-heavy) | Opus 4.7 | SWE-bench Pro +6.6, MCP-Atlas +10.1 vs GPT-5.4, 14% fewer tool errors vs Opus 4.6 |
| Short-form code completion / throughput | Sonnet 4.6 or GPT-5.4 Codex | Sonnet handles 90% of coding at 40% of Opus’s cost |
| Web research agents | GPT-5.4 Standard (budget) or Pro (accuracy) | Pro’s 89.3% BrowseComp is 10 pts ahead of Opus |
| Long-context (>200K tokens) | Claude Opus or Sonnet | Only lab with flat pricing across the full 1M window |
| Vision / multimodal | Opus 4.7 or Gemini 3.1 Pro | Opus vision enhanced to 2,576px; Gemini MMMU-Pro 80.5% |
| Math / STEM reasoning | Opus 4.7 or Gemini 3.1 Pro | GPQA Diamond: Opus 94.2%, Gemini 94.3%, GPT-5.4 92.8% — Opus and Gemini are effectively tied at the top |
| High-volume / budget | Haiku 4.5 or Gemini Flash-Lite | $1/$5 or $0.25/$1.50 — order of magnitude cheaper |
Real teams are already running this split. Per The New Stack’s AI coding tool analysis: Claude Code handles roughly 4% of all public GitHub commits (projected 20% by year-end per SemiAnalysis), while OpenAI Codex grew from 2M to 3M weekly active users in a single month. “Claude performs better on nuanced reasoning across long context windows, while Codex handles parallelizable throughput tasks more efficiently,” per The New Stack’s practitioner summary. Most teams use both.
The uncomfortable conclusion
Anthropic itself admits, in its Opus 4.7 launch announcement, that Opus 4.7 “is not a model that sweeps every leaderboard.” When the lab holding the flagship coding wins refuses to claim universal dominance, the question stops being “which model is best?” and becomes “how do you slot all three?”
For most production coding teams, Sonnet 4.6 is the smarter default. At $180/month versus $405/month for an identical coding agent workload — and with Anthropic’s own guidance saying Sonnet covers 90%+ of coding tasks — the burden of proof sits on Opus 4.7 to justify the premium on your specific traffic. The upgrade is a decision, not a default.
The next Artificial Analysis Intelligence Index quarterly update lands in late June 2026. That release will be the first read on whether the three-way 57-point tie holds or whether anyone breaks away as production teams push into longer agentic workflows and tokenizer inflation data matures. Until then, one open question decides your bill: if Sonnet 4.6 really handles 90%+ of coding tasks, what is the production case that justifies Opus 4.7 on your traffic — and does the tokenizer math ever close that gap, or permanently widen it?
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
