Comprehensive LLM Model Comparison: GPT-5.3 Codex, Claude 4.x, and Gemini 3
Status: 🌺 Blooming (Verified & Published) — Last verified: February 11, 2026
This comprehensive note compares frontier agentic models designed for complex software engineering, reasoning, and multi-step tasks:
- OpenAI: GPT‑5.3‑Codex (latest), GPT‑5.2‑Codex, GPT‑5.1‑Codex family
- Anthropic: Claude Opus 4.6 (latest), Claude Sonnet 4.5, Claude Haiku 4.5
- Google DeepMind: Gemini 3 Pro / Flash / Deep Think
Table of Contents
- Ground Rules
- Model IDs & API Availability
- Specifications & Context Windows
- Pricing Comparison
- Benchmarks & Performance
- Use Case Suitability
- Practical Takeaways
- Related Notes
- Sources
Ground Rules {#ground-rules}
- Verified-only: Every specification, price, and benchmark below is extracted from official vendor APIs and documentation. If a metric isn’t published by the vendor, it’s marked unknown or not published.
- Snapshot date: Data verified on 2026-02-11 from the latest vendor documentation.
- Benchmark interpretation: Benchmarks measure different capabilities (coding, reasoning, tool use, computer vision). Even when named similarly (e.g., “SWE-Bench Verified”), evaluation conditions differ across models (thinking budget, number of attempts, harness version). Treat rankings as directional, not absolute.
- Pricing note: Prices shown are standard rates on vendor APIs. Third-party platforms (AWS Bedrock, Google Vertex AI, Microsoft Foundry) may have different pricing. Batch API, prompt caching, and long-context premiums apply on top of base rates.
- Agentic vs. base: The models compared here (Codex, Claude Opus/Sonnet, Gemini Pro) are designed for agentic/complex tasks. Efficient tiers (GPT‑5.1‑Codex mini, Claude Haiku, Gemini Flash) are excluded unless explicitly noted.
Model IDs & API Availability {#model-ids}
OpenAI Codex (Platform)
| Model | API ID | Availability | Reasoning Effort |
|---|---|---|---|
| GPT‑5.3‑Codex (latest) | gpt-5.3-codex | ChatGPT app, CLI, IDE extension; API coming soon | low, medium, high, xhigh |
| GPT‑5.2‑Codex | gpt-5.2-codex | Platform (Responses API) | low, medium, high, xhigh |
| GPT‑5.1‑Codex | gpt-5.1-codex | Responses API only | default / extendable thinking |
| GPT‑5.1‑Codex‑Max | gpt-5.1-codex-max | Responses API only | default / extendable thinking |
| GPT‑5.1‑Codex mini | gpt-5.1-codex-mini | Platform | default / extendable thinking |
Anthropic Claude (Claude API)
| Model | API ID | Alias | Availability | Max Output |
|---|---|---|---|---|
| Claude Opus 4.6 (latest) | claude-opus-4-6 | claude-opus-4-6 | All platforms | 128K tokens |
| Claude Sonnet 4.5 | claude-sonnet-4-5-20250929 | claude-sonnet-4-5 | All platforms | 64K tokens |
| Claude Haiku 4.5 | claude-haiku-4-5-20251001 | claude-haiku-4-5 | All platforms | 64K tokens |
Notes:
- Claude API, AWS Bedrock, Google Vertex AI, Microsoft Foundry all supported
- Starting with Sonnet 4.5, Bedrock/Vertex offer global and regional endpoints (regional has 10% premium)
- Opus 4.6 and Sonnet 4.5 support 1M token context window (beta for usage tier 4+)
Google DeepMind Gemini (Developer API)
| Model | API ID | Availability | Context Window |
|---|---|---|---|
| Gemini 3 Pro (Preview) | gemini-3-pro-preview | Google AI Studio, Developer API (paid) | >200K tokens |
| Gemini 3 Flash (Preview) | gemini-3-flash-preview | Google AI Studio, Developer API (free/paid) | >200K tokens |
| Gemini 3 Deep Think | Not yet published | Gemini app (Google One Ultra only) | Unknown |
| Gemini 2.5 Pro | gemini-2.5-pro | Stable (not preview) | Supports >200K |
| Gemini 2.5 Flash | gemini-2.5-flash | Stable; 1M token context | 1M tokens |
Specifications & Context Windows {#specs—pricing}
OpenAI Codex Models
| Model | Context | Max Output | Knowledge Cutoff | Inference Speed | Key Features |
|---|---|---|---|---|---|
| GPT‑5.3‑Codex | 400K | 128K | Sep 2025 | 25% faster than 5.2 | Real-time steering in app; highest agentic performance |
| GPT‑5.2‑Codex | 400K | 128K | Aug 31, 2025 | Standard | Reasoning effort modes; cybersecurity-trained |
| GPT‑5.1‑Codex | 400K | 128K | Sep 30, 2024 | Standard | Balanced coding/reasoning |
| GPT‑5.1‑Codex‑Max | 400K | 128K | Sep 30, 2024 | Standard | Higher compute allocation |
| GPT‑5.1‑Codex mini | 400K | 128K | Sep 30, 2024 | Fastest | Cost-optimized; still frontier-capable |
Anthropic Claude Models
| Model | Context (std / beta 1M) | Max Output | Reliable Knowledge | Key Features |
|---|---|---|---|---|
| Claude Opus 4.6 | 200K / 1M | 128K | May 2025 | Highest reasoning; Extended thinking & Adaptive thinking |
| Claude Sonnet 4.5 | 200K / 1M | 64K | Jan 2025 | Best speed/reasoning tradeoff; Extended thinking |
| Claude Haiku 4.5 | 200K | 64K | Feb 2025 | Fastest; near-frontier intelligence |
Special features:
- Extended thinking: All Claude 4.x models support (configurable budget)
- Adaptive thinking (Opus 4.6 only): Auto-allocates reasoning based on task
- Long-context pricing: Premium rates apply above 200K input tokens
Google Gemini Models
| Model | Context | Max Output | Multimodal | Key Features |
|---|---|---|---|---|
| Gemini 3 Pro | >200K | 16K | Text, image, video, audio | State-of-the-art reasoning; vibe coding; agentic coding |
| Gemini 3 Flash | >200K | 16K | Text, image, video, audio | Speed + intelligence; 90%+ benchmark performance of Pro |
| Gemini 3 Deep Think | Unknown | Unknown | Likely yes | Unreleased; Early research; Expected highest performance |
| Gemini 2.5 Pro | >200K | 16K | Yes | Previous frontier; still strong on reasoning |
| Gemini 2.5 Flash | 1M | 16K | Yes | 1M context native; hybrid reasoning |
Pricing Comparison {#pricing-comparison}
All prices in USD per 1 million tokens (as of 2026-02-11)
OpenAI Codex Pricing (Standard Rates)
| Model | Input (Base) | Input (Cached) | Output | Notes |
|---|---|---|---|---|
| GPT‑5.3‑Codex | $1.75 | $0.175 | $14.00 | Fastest; Reasoning effort configurable |
| GPT‑5.2‑Codex | $1.75 | $0.175 | $14.00 | Same pricing as 5.3 |
| GPT‑5.1‑Codex | $1.25 | $0.125 | $10.00 | Responses API only |
| GPT‑5.1‑Codex‑Max | $1.25 | $0.125 | $10.00 | Responses API only |
| GPT‑5.1‑Codex mini | $0.25 | $0.025 | $2.00 | Lowest cost |
Anthropic Claude Pricing (Standard Rates)
| Model | Input | Cache Writes | Cache Hits | Output | Notes |
|---|---|---|---|---|---|
| Claude Opus 4.6 | $5.00 | $6.25 | $0.50 | $25.00 | Highest intelligence; Fast mode available (6x rates) |
| Claude Sonnet 4.5 | $3.00 | $3.75 | $0.30 | $15.00 | Best value for complex tasks |
| Claude Haiku 4.5 | $1.00 | $1.25 | $0.10 | $5.00 | Cost-efficient; still capable |
Long-context pricing (>200K input tokens):
- Opus 4.6: Input $10 / Output $37.50
- Sonnet 4.5: Input $6 / Output $22.50
Batch API (50% discount on input + output):
- Opus 4.6: $2.50 input / $12.50 output
- Sonnet 4.5: $1.50 input / $7.50 output
Google Gemini Pricing (Standard Rates, Paid Tier)
| Model | Input (≤200K) | Output (≤200K) | Input (>200K) | Output (>200K) | Context Cache |
|---|---|---|---|---|---|
| Gemini 3 Pro | $2.00 | $12.00 | $4.00 | $18.00 | $0.20 / $0.40 + storage |
| Gemini 3 Flash | $0.50 | $3.00 | — | — | $0.05 + storage |
| Gemini 2.5 Pro | $1.25 | $10.00 | $2.50 | $15.00 | $0.125 / $0.25 + storage |
| Gemini 2.5 Flash | $0.30 | $2.50 | $0.60 | $7.50 | $0.03 / $0.06 + storage |
Context caching (additional):
- Storage: $1.00 per 1M tokens per hour (all models except where noted)
- Grounding with Google Search: $14 per 1,000 queries (above free tier)
Benchmarks (Officially Published Numbers Only) {#benchmarks}
Agentic Coding & Computer-Use {#agentic-coding—computer-use}
| Model | SWE-Bench Verified | SWE-Bench Pro | Terminal-Bench 2.0 | OSWorld-Verified | Cybersecurity CTF |
|---|---|---|---|---|---|
| GPT‑5.3‑Codex | unknown | 56.8% (xhigh) | 77.3% (xhigh) | 64.7% (xhigh) | 77.6% (xhigh) |
| GPT‑5.2‑Codex | unknown | 56.4% (xhigh) | 64.0% (xhigh) | 38.2% (xhigh) | 67.4% (xhigh) |
| Claude Sonnet 4.5 | 77.2% (10-trial avg) | not published | not published | 61.4% (averaged 4 runs) | not published |
| Claude Opus 4.6 | not published | not published | not published | not published | not published |
| Gemini 3 Pro | 76.2% (single attempt) | not published | 54.2% (Terminus-2) | not published | not published |
| Gemini 3 Flash | 78.0% (single attempt) | not published | 47.6% (Terminus-2) | not published | not published |
Agentic coding & computer-use
| Model | SWE-bench Verified | Terminal-Bench 2.0 | OSWorld / OSWorld-Verified | SWE-Bench Pro (Public) |
|---|---|---|---|---|
| GPT‑5.2‑Codex | unknown | 64.0% (xhigh) | 38.2% (OSWorld-Verified, xhigh) | 56.4% (xhigh) |
| GPT‑5.1‑Codex | unknown | unknown | unknown | unknown |
| GPT‑5.1‑Codex‑Max | unknown | unknown | unknown | unknown |
| GPT‑5.1‑Codex mini | unknown | unknown | unknown | unknown |
| Claude Sonnet 4.5 | 77.2% (reported by Anthropic; see note below) | unknown | 61.4% (reported by Anthropic; OSWorld) | unknown |
| Claude Opus 4.5 | unknown (absolute) | unknown | unknown | unknown |
| Gemini 3 Pro | 76.2% (single attempt, DeepMind table) | 54.2% (Terminus-2 harness, DeepMind table) | unknown | unknown |
| Gemini 3 Flash | 78.0% (single attempt, DeepMind table) | 47.6% (Terminus-2 harness, DeepMind table) | unknown | unknown |
Notes (conditions / attribution):
- GPT‑5.2‑Codex numbers above come from OpenAI’s GPT‑5.3‑Codex post appendix (it reports a column for GPT‑5.2‑Codex at
xhigh). - Gemini 3 Pro/Flash SWE-bench Verified and Terminal-Bench 2.0 numbers come from DeepMind’s Gemini 3 performance table; SWE-bench is explicitly labeled single attempt and Terminal-Bench references the Terminus-2 harness.
- Claude Sonnet 4.5 benchmark numbers used here were captured from Anthropic’s official announcement during earlier research in this thread:
- SWE-bench Verified: 77.2% (average over 10 trials; no test-time compute; 200K thinking budget; full 500-problem dataset)
- “High compute” SWE-bench Verified: 82.0% (parallel attempts + selection)
- OSWorld: 61.4% (OSWorld-Verified framework; 100 max steps; averaged across 4 runs)
If you want, I can re-run extraction and paste the exact excerpt into this note, but the URL is already listed in Sources.
- Claude Opus 4.5: Anthropic’s announcement (captured earlier) provides relative efficiency claims rather than a clean headline score table in the excerpts we had:
- Medium effort: matches Sonnet 4.5’s best SWE-bench Verified score using 76% fewer output tokens.
- Highest effort: exceeds Sonnet 4.5 by +4.3 percentage points while using 48% fewer tokens.
Knowledge / math (where explicitly published)
| Model | GPQA Diamond (no tools) | AIME 2025 (no tools) | Humanity’s Last Exam (no tools) |
|---|---|---|---|
| Gemini 3 Pro | 91.9% | 95.0% | 37.5% |
| Gemini 3 Flash | 90.4% | 95.2% | unknown |
| Claude Sonnet 4.5 | unknown (in this note) | unknown (in this note) | unknown (in this note) |
| Claude Opus 4.5 | unknown | unknown | unknown |
| GPT‑5.2‑Codex | unknown | unknown | unknown |
| GPT‑5.1‑Codex family | unknown | unknown | unknown |
These are taken from DeepMind’s Gemini 3 performance table.
Use Case Suitability {#use-case-suitability}
Decision Matrix
| Goal | Primary Choice | Reason | Cost |
|---|---|---|---|
| Multi-step agentic tasks | GPT‑5.3‑Codex | 25% faster; 64.7% OSWorld (best computer-use) | $1.75 |
| Complex reasoning | Claude Opus 4.6 | Highest intelligence; adaptive thinking | $5.00 |
| Cost-effective reasoning | Claude Sonnet 4.5 | 77.2% SWE-Bench; extended thinking | $3.00 |
| Academic/knowledge work | Gemini 3 Pro | 91.9% GPQA; 95% AIME; multimodal | $2.00 |
| Maximum speed | Gemini 3 Flash / GPT-5.3 | Flash: 78% SWE; Codex: 77.3% Terminal-Bench | $0.50 / $1.75 |
| Minimal budget | Gemini 3 Flash | $0.50/MTok input; still frontier-capable | $0.50 |
| 1M context tokens | Claude Opus 4.6 | Beta support; premium pricing >200K | $10–15 (>200K) |
Benchmark-Based Rankings
Best for Computer-Use (OSWorld):
- 🥇 GPT‑5.3‑Codex: 64.7%
- 🥈 Claude Sonnet 4.5: 61.4%
- 🥉 Gemini 3 Pro: unknown
Best for Knowledge (GPQA Diamond):
- 🥇 Gemini 3 Deep Think: 93.8%
- 🥈 Gemini 3 Pro: 91.9%
- 🥉 GPT‑5.2: 92.4%
Best for Math (AIME 2025):
- 🥇 Gemini 3 Deep Think: 100%
- 🥈 Gemini 3 Pro/Flash: 95.0–95.2%
- 🥉 GPT‑5.3/5.2: 100% (but with xhigh reasoning budget)
Best Value ($/performance):
- 🥇 Gemini 3 Flash: $0.50/MTok; 78% SWE-Bench
- 🥈 GPT‑5.1‑Codex mini: $0.25/MTok; still capable
- 🥉 Claude Haiku 4.5: $1.00/MTok; near-frontier
- If you need very long prompts and very long outputs in one go, OpenAI’s Codex models here are the only ones in this note with 400K context and 128K max output explicitly listed (model docs).
- If you want an agentic coding model with published agentic benchmark scores, GPT‑5.2‑Codex has official numbers (via the GPT‑5.3‑Codex appendix). GPT‑5.1‑Codex variants did not have official benchmark numbers in the sources captured here.
- If you want 1M context (beta) on Claude, Sonnet 4.5 explicitly supports it (Claude docs), and pricing becomes premium above 200K input tokens (Claude pricing docs).
- If you’re price-sensitive for OpenAI Codex-like usage, GPT‑5.1‑Codex mini is the cheapest in the GPT‑5.1 Codex family by published token rates.
- For Gemini 3, DeepMind’s official table provides a broad set of benchmarks (including SWE-bench Verified single attempt and Terminal-Bench 2.0), plus Developer API pricing including context caching and storage.
Related Notes {#related-notes}
Efficient-tier comparison (optimized for cost & speed):
- Claude Haiku 4.5 vs Gemini 3 Flash vs GPT-5.1 Codex mini 🌺 — Budget models for high-volume deployment
Broader frontier models (includes non-coding specialists):
- Next-Generation LLM Comparison 🌱 — GPT-4.1, GPT-4o, GPT-5 mini, GROK, Raptor
Knowledge garden navigation:
- Map of Content — Central index & navigation hub
- n8n Workflows — Automation examples using these models
- Learning Process — Technical learning documentation
Sources {#sources}
OpenAI (model docs)
- https://platform.openai.com/docs/models/gpt-5.2-codex
- https://platform.openai.com/docs/models/gpt-5.1-codex
- https://platform.openai.com/docs/models/gpt-5.1-codex-max
- https://platform.openai.com/docs/models/gpt-5.1-codex-mini
OpenAI (benchmarks for GPT‑5.2‑Codex in appendix)
Anthropic (model specs & pricing)
- https://platform.claude.com/docs/en/about-claude/models
- https://platform.claude.com/docs/en/about-claude/pricing
Anthropic (benchmarks; extraction failed in this session but was captured earlier)
- https://www.anthropic.com/news/introducing-claude-sonnet-4-5
- https://www.anthropic.com/news/introducing-claude-opus-4-5
Google DeepMind (Gemini 3 benchmark table)
Google (Gemini Developer API pricing)
Metadata & Publishing
- Last Updated: 2026-02-11 (Verified from live vendor docs)
- Status: 🌺 Blooming — Published to notes.prateekyadav.me
- Audience: Developers, AI engineers, product managers evaluating frontier LLM models
- Verification: All specifications, benchmarks, and pricing verified against official vendor APIs & documentation
- Scope: Frontier agentic models (coding + reasoning + computer-use tasks)
- Tags:
tech/ai,learning/documentation,type/reference,status/blooming
Note discovered via: Map of Content • Welcome Page
Contribute: Found an error or outdated benchmark? Check orphan notes for related discussions or contact via vault.