LLM Model Scorecard: GPT-5, Claude 4, and Gemini 3 compared side-by-side ## TL;DR There is no single “best” frontier model in 2026. The right choice depends on workload: reasoning depth, cost tolerance, latency requirements, tool-use reliability, data governance needs, and multimodal support. Based on current evidence (checked 2026-05-30): | Dimension | Strongest performer | Notes |
| | | | | Reasoning & knowledge | Claude 4 Opus / GPT-5 premium | Tight race; workload-specific evaluation needed | | Coding & tool use | GPT-5 / Claude Sonnet 4.5 | Both strong; GPT-5 edges ahead on structured tool call reliability | | Cost efficiency | Gemini 2.5 Pro / GPT-5 mid-tier | Gemini wins at high cache-hit rates; GPT-5 mid-tier has competitive per-task cost | | Context handling | Gemini 2.5 Pro (2M) | 2M context window; others at 200K–1M | | Multimodal | Gemini (native audio/video) | Others strong on images; Gemini leads on native audio and video input | | Speed/latency | Gemini 2.0 Flash / GPT-5 fast | Budget-speed tiers from both providers | | Data governance | Varies by provider tier | Enterprise vs API-default terms differ significantly — see Data Governance section | The safe approach: shortlist two models from different providers, run a 50–100 case evaluation on your actual workload, and compare cost per completed task — not cost per token or leaderboard rank.
## The frontier scorecard table The table below compares GPT-5, Claude 4, and Gemini 3 across the dimensions that matter for procurement. Each row links to the site's deeper article on that topic. ### Model availability & context | Dimension | OpenAI GPT-5 | Anthropic Claude 4 | Google Gemini 2.5 | | | | | | | **Current flagship models** | GPT-5 (premium tier), GPT-5 mini (mid-tier), GPT-5 fast (budget) | Claude Opus (premium), Claude Sonnet 4.5 (workhorse), Claude Haiku 3.5 (budget) | Gemini 2.5 Pro (premium), Gemini 2.0 Flash (budget) | | **Max context window** | 200K tokens (GPT-5 premium); 128K (mini/fast) | 200K tokens (all models) | 2M tokens (Pro); 1M tokens (Flash) | | **Training data cutoff** | Unpublished (varies by model snapshot) | Early 2026 (Opus/Sonnet) | Late 2025 (Pro/Flash) | | **Knowledge recency** | Weekly-to-daily updates for some tiers | Periodic model snapshot updates | Regular model refresh cadence | | **Structured outputs** | ✅ JSON mode + structured outputs | ✅ JSON mode + structured outputs (tool-use) | ✅ JSON mode + structured outputs | | See: [Context windows explained](/cache/context-windows-explained-why-bigger-is-not-always-better/) | | | | ### Quality benchmarks (approximate standings, 2026-05-30) | Benchmark | GPT-5 | Claude 4 | Gemini 2.5 | Notes | | | | | | | | **MMLU (professional)** | ~92% | ~91% | ~90% | Within margin of methodology noise | | **LiveBench (overall)** | Top cluster | Top cluster | Top cluster | All three in statistical tie; see LiveBench weekly | | **Chatbot Arena ELO** | ~1380 | ~1360 | ~1350 | Preference-based; varies week to week | | **HumanEval / coding** | ~88% | ~85% | ~84% | GPT-5 edges ahead on structured coding tasks | | **SWE-bench (real-world coding)** | ~55% | ~52% | ~48% | GPT-5 leads on full-repo coding tasks | | **Long-context recall** | Strong (200K needle tests) | Strong (200K needle tests) | Very strong (2M documented recall) | Gemini benefits from larger context window | | **Tool-use reliability** | ~92% function call accuracy | ~90% function call accuracy | ~85% function call accuracy | GPT-5 leads on reliable structured tool calls | | See: [Benchmark leaderboards](/cache/benchmark-leaderboards-chatbot-arena-livebench/) · [How LLM benchmarks work](/evals/how-llm-benchmarks-work-and-what-they-miss/) | | | | **Benchmark caution:** These scores are approximate and date-stamped. Benchmark methodology, contamination and saturation affect all providers. A 2–3 point gap is rarely meaningful for procurement — treat these as shortlist filters, not rankings. ### Pricing snapshot (USD per million tokens, 2026-05-30) | Provider / Model | Input ($/M) | Cached input ($/M) | Output ($/M) | Context | Batch discount | | | | | | | | | **OpenAI GPT-5 mini** | ~$2.00* | ~$1.00* | ~$10.00* | 128K | 50% batch | | **OpenAI GPT-5 (premium)** | ~$10.00* | ~$5.00* | ~$50.00* | 200K | 50% batch | | **Claude Haiku 3.5** | ~$0.80 | ~$0.08 (cache hit) | ~$4.00 | 200K | Possible† | | **Claude Sonnet 4.5** | $3.00 | $0.30 (cache hit) | $15.00 | 200K | Possible† | | **Claude Opus** | $15.00 | ~$1.50 (cache hit) | ~$75.00 | 200K | Possible† | | **Gemini 2.0 Flash** | $0.15 | $0.075 | $0.60 | 1M | 50% batch | | **Gemini 2.5 Pro (≤200K)** | $1.25 | $0.13 | $10.00 | 2M | 50% batch | | **Gemini 2.5 Pro (>200K)** | $2.50 | $0.25 | $15.00 | 2M | 50% batch | *OpenAI rates from third-party aggregator (Artificial Analysis) — OpenAI pricing pages blocked server-side access. Verify directly at platform.openai.com. †Anthropic batch FAQ mentions cache pricing can combine with batch discounts but does not publish a simple batch row price. The full pricing deep-dive with worked cost examples is at [LLM API pricing comparison 2026](/cache/llm-api-pricing-comparison-2026/). **Key pricing insight:** Budget models (Gemini Flash, GPT-5 fast tier, Claude Haiku) are within 2–5× of each other for simple tasks. Premium models spread by 5–10×. Cache economics — not headline rates — often determine the real cost winner at scale. ### Speed & latency characteristics | Dimension | GPT-5 | Claude 4 | Gemini 2.5 | | | | | | | **Time to first token (TTFT)** | Fast (~200–500ms) | Moderate (~500–1500ms) | Fast (~200–400ms) | | **Output throughput** | ~40–60 tok/s (mini), ~20–40 tok/s (premium) | ~30–50 tok/s (Haiku), ~15–25 tok/s (Sonnet/Opus) | ~50–80 tok/s (Flash), ~20–40 tok/s (Pro) | | **Prompt caching benefit** | ~50% discount (cached input) | ~90% discount (cache hit); cache writes cost premium | ~90% discount (≤200K), ~90% (>200K) | | **Batch turnaround** | 4–24 hours | Hours to same-day | Hours to same-day | | **Rate limit tiers** | Free/paid tiers; overage possible | Published rate limits per tier | Published rate limits per tier | | See: [Latency in LLM apps](/run/latency-in-llm-apps-first-token-total-time-and-user-experience/) · [Prompt caching](/cache/prompt-caching-explained-when-repeated-context-becomes-cheaper/) | | | | ### Modality & capability support | Capability | GPT-5 | Claude 4 | Gemini 2.5 | | | | | | | **Text input** | ✅ | ✅ | ✅ | | **Image input (understanding)** | ✅ | ✅ | ✅ | | **Audio input** | ✅ (Whisper-based) | ❌ (text/image only) | ✅ (native audio) | | **Video input** | ✅ | ❌ (can process video frames as images) | ✅ (native video) | | **Document understanding** | ✅ (PDF, images) | ✅ (PDF, images) | ✅ (PDF, images, native) | | **Code execution / sandbox** | ✅ (Code Interpreter) | ✅ (Analysis tool) | ✅ (Code execution) | | **Function calling / tool use** | ✅ (strong) | ✅ (strong) | ✅ (good) | | **Multilingual** | ✅ (broad) | ✅ (broad) | ✅ (broad, strongest non-English) | | **Structured output (JSON mode)** | ✅ | ✅ | ✅ | | **Vision reasoning** | ✅ | ✅ | ✅ (native spatial reasoning) | | See: [Multimodal models explained](/cache/multimodal-models-explained-text-images-audio-and-video-in-practical-products/) | | | | **Multimodal note:** All three providers can "see" images and extract information from them. Gemini is the only provider offering native audio and video input without a separate transcription or frame-extraction pipeline, which matters if your workload processes recorded meetings, live streams, or phone calls directly. ### Data governance & portability | Dimension | OpenAI | Anthropic | Google | | | | | | | **Training data use (API)** | Opt-out available | Opt-out available (default off for paid API) | Opt-out available | | **Data retention (default)** | 30 days (API) | 30 days (API) | Varies by service tier | | **Region controls** | US and EU available | US and EU available (EU data residency) | Multi-region via GCP | | **Model versioning** | Model aliases can change; pinned versions available | Model snapshots documented; version aliases | Version aliases documented | | **SOC 2 / ISO 27001** | ✅ | ✅ | ✅ | | **API logs for monitoring** | Available (retention configurable) | Available (30-day default configurable) | Available (via GCP) | | **Export options** | Logs export available | Logs export available | Logs export via GCP | | See: [Provider data retention policies](/diff/provider-data-retention-policies-what-api-users-should-compare/) · [AI vendor lock-in](/diff/ai-vendor-lock-in-model-apis-embeddings-vector-stores-and-eval-data/) | | | | ## How to use this scorecard This comparison table is designed as a starting point for your own evaluation. Do not pick a model from this page alone. ### Step 1: Shortlist Use the scorecard to narrow from all available models to 2–3 candidates that fit your workload's primary constraint. For example: - **Cost-sensitive, high-volume classification:** Shortlist GPT-5 mini, Gemini 2.0 Flash, Claude Haiku — budget models first. - **Complex reasoning with customer-facing output:** Shortlist Claude Opus, GPT-5 premium — quality over cost. - **Multimodal processing (audio/video):** Shortlist Gemini 2.5 Pro only, then verify against alternatives. - **High-reliability tool use (agent workflows):** Shortlist GPT-5 and Claude Sonnet 4.5 — both strong on function calling. ### Step 2: Evaluate Build a 50–100 case test set from your actual workload. Include: - Common cases that represent 80% of your traffic. - Edge cases that have caused failures before. - Known-answer questions to check factual grounding. - Safety-relevant or policy-sensitive boundaries. Run the same prompts through each shortlisted model. Score outputs against the same criteria. Compare cost per completed task, not cost per token. The guide [Creating a model scorecard for your own workload](/cache/creating-a-model-scorecard-for-your-own-workload/) provides a repeatable framework for this step. ### Step 3: Decide Pick the model that passes your evaluation threshold at the lowest cost per completed task. Plan for: - A fallback model from a different provider (for redundancy, not just quality). - A re-evaluation schedule — every 90 days at minimum, more often if pricing or model versions change. - An exit path — how would you switch if your chosen provider changes terms, pricing, or availability? ## What this scorecard does not cover - **Self-hosted / open-weight models:** Llama 4, DeepSeek V4, Qwen, Mistral and others are outside this comparison. See [Hosted API vs self-hosted open model](/comparisons/hosted-api-vs-self-hosted-open-model/) for that choice. - **Fine-tuned or custom models:** Provider-specific fine-tuning options are not compared here. See [Fine-tuning economics](/comparisons/fine-tuning-economics/). - **Model gateways / routers:** OpenRouter, LiteLLM and cloud AI platforms (Bedrock, Vertex, Azure) are separate decisions. See [Model gateways compared](/comparisons/model-gateways-and-routers-openrouter-litellm-and-build-vs-buy/). - **Small language models:** Models under 7B parameters are not included. See [Small language models](/diff/small-language-models-when-smaller-is-better/). - **Regulatory compliance per jurisdiction:** This comparison covers general data governance. For UK, EU or US-specific regulatory considerations, see [LLMs in regulated industries](/diff/llms-in-regulated-industries-compliance-infrastructure-and-deployment-patterns/). ## When to check again The ratings in this scorecard should be revisited when: 1. A provider releases a new flagship model or significant update. 2. Pricing changes by more than 20% on any model tier. 3. A new benchmark or evaluation methodology shifts the relative standings. 4. Your workload changes — new task types, new data sensitivity requirements, new volume projections. 5. It has been more than 90 days since the last check. ## Methodology - **Data checked:** 2026-05-30 - **Sources consulted:** Provider documentation, model cards, pricing pages, status pages and benchmark leaderboards from OpenAI, Anthropic and Google (Gemini). Third-party aggregators cited where direct access was blocked or unreliable. - **Directly verified at source (HTTP 200 from this server):** Anthropic, Google, DeepSeek (where referenced) - **Blocked from server-side access:** OpenAI pricing pages (HTTP 403) — rates estimated from third-party aggregator Artificial Analysis - **Benchmark scores:** Approximate cluster positions, not precise ranks. Methodology noise, contamination and saturation make sub-5% differences unreliable for procurement. - **Assumptions:** Pricing assumes standard pay-as-you-go API access without enterprise volume discounts or reserved capacity. Latency figures are provider-published or third-party measured ranges on paid tiers. Benchmark scores are drawn from the most recent published results as of the check date. - **Limitations:** This scorecard provides a comparison snapshot, not live intelligence. Model versions, pricing and benchmark positions change frequently. It does not cover fine-tuned models, self-hosted inference, or provider-specific enterprise features. It is not legal, financial or procurement advice. - **Jurisdiction:** Global. Data governance references are drawn from each provider's public documentation. Local requirements (GDPR, CCPA, UK AI regulation) may impose additional constraints not fully covered here. ## Source list - [OpenAI models overview](https://platform.openai.com/docs/models) — accessed 2026-05-30 - [OpenAI API pricing](https://openai.com/api/pricing/) — inaccessible from this server; cross-referenced via Artificial Analysis - [Anthropic Claude models](https://docs.anthropic.com/en/docs/about-claude/models) — accessed 2026-05-30 - [Anthropic pricing](https://www.anthropic.com/pricing) — accessed 2026-05-30 - [Google Gemini API models](https://ai.google.dev/gemini-api/docs/models) — accessed 2026-05-30 - [Google Gemini API pricing](https://ai.google.dev/pricing) — accessed 2026-05-30 - [Artificial Analysis model comparison](https://artificialanalysis.ai/models) — accessed 2026-05-30 - [LMSYS Chatbot Arena](https://lmarena.ai/) — accessed 2026-05-30 - [LiveBench](https://livebench.ai/) — accessed 2026-05-30 - [SWE-bench](https://www.swebench.com/) — accessed 2026-05-30 - [OpenAI data controls FAQ](https://help.openai.com/en/articles/7039943-data-controls-faq) — accessed 2026-05-30 - [Anthropic privacy policy](https://www.anthropic.com/legal/privacy) — accessed 2026-05-30 - [Google Cloud data processing](https://cloud.google.com/terms/data-processing-addendum) — accessed 2026-05-30 ## Related guides - [Creating a model scorecard for your own workload](/cache/creating-a-model-scorecard-for-your-own-workload/) — the process guide for building your own evaluation - [LLM API pricing comparison 2026](/cache/llm-api-pricing-comparison-2026/) — deep-dive on pricing with worked cost examples - [Benchmark leaderboards for busy buyers](/cache/benchmark-leaderboards-chatbot-arena-livebench/) — how to read leaderboards without being misled - [OpenAI, Anthropic, Google and Mistral APIs: what comparison pages should measure](/diff/openai-anthropic-google-and-mistral-apis-what-comparison-pages-should-measure/) — the rubric for comparing providers - [How LLM benchmarks work, and what they miss](/evals/how-llm-benchmarks-work-and-what-they-miss/) — understanding what benchmark scores actually measure - [Multimodal models explained](/cache/multimodal-models-explained-text-images-audio-and-video-in-practical-products/) — understanding modality differences across providers - [Provider data retention policies](/diff/provider-data-retention-policies-what-api-users-should-compare/) — data governance comparison - [Latency in LLM apps](/run/latency-in-llm-apps-first-token-total-time-and-user-experience/) — understanding TTFT, throughput and user experience ## Trust Stack - **Last checked:** 2026-05-30 - **Corrections:** [Contact us](/contact/) to report errors ## Change log - **2026-05-30:** First draft. Side-by-side frontier model scorecard for GPT-5, Claude 4, and Gemini 3 covering benchmarks, pricing, context, modalities, speed and data governance.