ChatGPT vs Claude vs Gemini in 2026: Identical Tasks, Honest Scores
Every other week, someone publishes a “Claude vs ChatGPT vs Gemini” article based on one prompt and a vibe. We’ve been doing this with rigor for six months. Same prompts, multiple runs, published rubric.
This article reflects our Q2 2026 LLM benchmark refresh with the current frontier models:
- Claude Opus 4.6 (Anthropic, via Claude.ai Pro)
- GPT-5 (OpenAI, via ChatGPT Plus)
- Gemini 3 Pro (Google, via Gemini Advanced)
We ran 24 tasks across 4 categories. Each task was prompted 5 times to control for variance. The full rubric is published — you can re-weight to your own priorities.
Headline scores (weighted total)
Out of 100:
| Model | Writing | Reasoning | Coding | Math | Weighted Total |
|---|---|---|---|---|---|
| Claude Opus 4.6 | 87 | 84 | 89 | 81 | 85.4 |
| GPT-5 | 82 | 88 | 85 | 89 | 86.1 |
| Gemini 3 Pro | 79 | 81 | 80 | 86 | 81.5 |
Verdict: GPT-5 narrowly edges out Claude on weighted total because of stronger reasoning and math. But the gap is small enough (0.7 points) that per-category strengths matter more for your decision.
By category:
– Writing-heavy workflows: Claude wins clearly
– Pure reasoning / abstract problems: GPT-5 wins
– Coding: Claude leads by 4 points
– Math (esp. competition-style): GPT-5 dominates
– Multimodal (image gen + voice): GPT-5 has the most integrated experience; Gemini close
What we tested
Writing (6 tasks)
- 1500-word essay from a brief
- Short story with constraints (genre, pov, length)
- Email rewrite for tone (formal → casual)
- Long-form article with specific structure
- Technical documentation
- Persuasive writing (op-ed style)
Scored on: prose quality, adherence to brief, structure, tone control, “doesn’t sound like AI” (judged blind by 3 human reviewers).
Reasoning (6 tasks)
- Multi-step logic puzzle
- Causal analysis of a scenario
- Identifying flaws in an argument
- Strategic recommendation given constraints
- Trade-off analysis between options
- Knowledge integration across domains
Scored on: correctness, completeness of reasoning chain, identifying edge cases.
Coding (6 tasks)
- Bug fix in a 200-line Python file (real bug from open-source)
- Implement a feature with provided spec
- Refactor a working but messy function
- Write tests for an untested function
- Explain a complex regex
- Translate between languages (Python ↔ TypeScript)
Scored on: correctness (does it run? does it pass tests?), code style, explanation quality, handling of edge cases.
Math (6 tasks)
- High school algebra word problem
- Calculus problem (definite integral with substitution)
- Probability problem (Bayesian)
- Statistics problem (interpreting confidence intervals)
- Competition-style number theory
- Applied math (financial / engineering)
Scored on: correct final answer, correct method, clarity of work shown.
Per-dimension scoring (1-5 scale)
We score every model on 6 dimensions on every task. Averaged across all 24 tasks:
| Dimension | Claude | GPT-5 | Gemini |
|---|---|---|---|
| Quality of output | 4.4 | 4.5 | 4.1 |
| Latency | 4.2 | 4.3 | 4.4 |
| Cost per task | 4.1 | 4.0 | 4.3 |
| Reliability (variance) | 4.5 | 4.2 | 4.0 |
| UX / ergonomics | 4.3 | 4.4 | 4.1 |
| Safety / accuracy | 4.6 | 4.2 | 4.3 |
Reliability is where Claude’s lead matters most. The same prompt produced equivalent-quality output across 5 runs more often for Claude than for GPT-5 or Gemini. If you can’t predict what you’ll get from your prompt, the model is less useful in production.
Safety / accuracy measures both refusal rates (do they refuse legitimate requests?) and hallucination rates (do they make stuff up?). Claude scored highest here — it refuses less than the others and hallucinates less in our factual probes.
Where each model is best
Claude Opus 4.6 — best for:
- Long-form writing that needs to sound human and maintain tone over 1000+ words
- Coherent multi-file code edits in real codebases
- Anything involving judgment calls (writing pushback, weighing trade-offs, analyzing nuance)
- Long context work (200K+ tokens with consistent quality)
- Workflows where reliability matters more than peak quality
If 60%+ of your LLM usage is writing or coding, Claude is the pick.
GPT-5 — best for:
- Pure reasoning without writing constraints (math problems, logic puzzles, strategic analysis)
- Multimodal workflows that combine image generation, voice mode, file analysis
- Use cases requiring tool use (function calling, integrated search, code execution)
- Math, especially competition-style or higher-level
- Quick factual Q&A with high accuracy
If your work mixes text, images, voice, and you want one integrated tool, GPT-5 (via ChatGPT Plus) is the pick.
Gemini 3 Pro — best for:
- Workflows already inside Google Workspace (Docs, Sheets, Gmail integration)
- Long-document analysis where Gemini’s 1M+ token context is genuinely usable
- Multilingual tasks especially in less-resourced languages
- Visual reasoning (Gemini’s vision capabilities are strong)
If you live in Google Workspace and need an LLM that works there, Gemini is the pick. Outside Workspace, it’s a reasonable but not standout choice.
What about cost?
Pro tier (Plus, Pro, Advanced) — all three are $20/mo for the basic consumer subscription. From a pure cost-per-task perspective:
- Claude Opus 4.6 via Claude.ai Pro: $20/mo, fair usage limits
- GPT-5 via ChatGPT Plus: $20/mo, fair usage limits, plus extras (image gen, voice mode, search)
- Gemini 3 Pro via Gemini Advanced: $20/mo, fair usage limits, plus Google One 2TB storage
If you’re using one model heavily, all three are similar value. ChatGPT Plus offers more bundled features (image gen, voice mode) so feels like better “per dollar” if you’d use those features.
For API access (developers, automation):
- Claude Opus 4.6 API: $15 in / $75 out per million tokens
- GPT-5 API: $10 in / $40 out per million tokens
- Gemini 3 Pro API: $7 in / $30 out per million tokens
Gemini is cheapest API. For high-volume automation where you don’t need peak quality, Gemini wins on cost.
What about Llama 4.5, Mistral, Qwen, DeepSeek?
We benchmarked the latest open-weight models alongside in this same testing window. The current frontier-open-weight models score ~75-80 on our weighted total — 5-10 points below the frontier closed models.
For automation workloads where you can host them yourself, the open-weight options are excellent (zero per-call cost, no rate limits, privacy benefits). For consumer use, the frontier closed models still win on raw quality.
We have a separate benchmark article for open-weight: Best Open-Weight LLMs 2026.
The recommendation tree
Use Claude if:
– Writing is 40%+ of your usage
– You value reliable, repeatable quality
– You want a “thinking partner” feel
– Your work involves nuanced judgment calls
– You need long context (codebases, long docs)
Use ChatGPT (GPT-5) if:
– Your work mixes text, images, voice, file analysis
– You want one integrated tool with the most features
– Math/reasoning is a meaningful part of your usage
– You’ll use the bundled image generation
– You want the most “default” assistant experience
Use Gemini if:
– You live in Google Workspace
– You need multilingual support in less-common languages
– You want the cheapest API for automation
– You’re already paying for Google One storage
What we’d actually do
If we could only have one: Claude Opus 4.6. The reliability score plus writing quality is what matters for daily use.
If we could have two: Claude + ChatGPT Plus. Claude as primary, ChatGPT for the image gen + voice + reasoning gaps.
If we could only spend $20/mo: Claude Pro. The other two are great but not enough better at their strengths to displace Claude’s daily-driver fit.
What’s coming Q3 2026
The benchmark will refresh quarterly. Already in our test queue:
– GPT-5.5 (rumored release)
– Claude 4.7 / 5.0 (expected refresh)
– Gemini 3 Ultra (released, not yet benchmarked)
– New open-weight competitors (Llama 5, Mistral Large 3)
Subscribers to AI Tools Tested get the new benchmark in the first week of each quarter.
Disclaimer & affiliate disclosure
OpenAI offers an affiliate program for ChatGPT. Anthropic does not currently offer a public Claude affiliate program. Google does not offer one for Gemini Advanced. We recommend Claude as our top pick despite earning zero commission on it — the benchmarks support it. See our affiliate disclosure.
This benchmark reflects model behavior at the time of testing (April-May 2026). LLM models update frequently and quietly — a model that scored X last month may behave differently this month. Re-verify before committing to a long-term subscription.
Last updated 2026 Q2. 5-run prompts across 24 tasks. Full rubric and per-task scores available on request.