GPT-5 accuracy reaches 92.6% on standard benchmarks
GPT-5 reaches 92.6% accuracy on standard benchmark tests. This is a 11.90% improvement over GPT-4o in factual answers, math problems, and coding tasks. The model’s hallucination rate is 12%. This rate is 45% lower than GPT-4o and 80% lower than the o3 model. GPT-5 reaches 94.6% on the AIME 2025 competition-level maths test. It also scores 88% on Aider Polyglot for multi-language code editing. The model achieves 46.2% on the HealthBench Hard evaluation. GPT-5 also scores 42% on Humanity’s Last Exam questions. While GPT-5 delivers clear accuracy gains on standard tests, beating GPT-4o by 11.90% in factual answers, math problems, and coding tasks, the model still produces inaccurate information that requires human oversight for verification and refinement. The system uses a real-time router to choose between fast replies and deep reasoning. This router decides based on task complexity or user intent.
GPT-5 is a unified system. It includes a smart base model, a reasoning mode called GPT-5 thinking, and a real-time router. The router chooses between fast replies and deeper analysis automatically. The model is smarter across the board. It provides more useful responses across math, science, finance, and law. GPT-5 also shows improvements in writing, including poetry, structural editing, and literary expression. It can generate complex front-end code and debug large codebases. It also maintains consistency across long-form content because it has 400,000 tokens of context. The model has an output window of 128,000 tokens.
GPT-5 errors in multimodal and logic tasks
Users found significant errors in the GPT-5 live demonstration. A Hacker News thread dissected the demo of the Bernoulli effect. The model gave an incorrect explanation of the phenomenon. It also struggled with visual comprehension in generative images. When a user asked to move a logo in a virtual background, the model placed the logo in the center instead of the corner. It also generated a fake microphone that obstructed the background. In professional tasks, GPT-5 incorrectly identified the number of controls in the CMMC Level 1 guidance. The model stated there were 17 controls. The actual number of controls is 15.
The model struggles with following the rules in certain domains. A Tufts professor found an example where GPT-5 became lost during a simple chess problem. The model also has issues with the challenge of parts and wholes in generative images. Some users say the model is too stiff and formal. The model often defaults to shorter, more abrupt responses. This is a departure from the conversational outputs of GPT-4. Many users also found that GPT-5 failed to provide correct answers for certain mathematical reasoning tasks. Does the model truly possess PhD level expertise in every area?
Claude 4.7 performance and technical specifications
Claude 4.7 is a reasoning model from Anthropic. It arrived on April 16, 2026. It reaches 95.8% on AIME 2026 Mathematics when using tools. The model scores 91.4% on GPQA Diamond. It reaches 87.6% on SWE-bench Verified. The model scores 70.2% on FrontierMath v2. It also achieves 79.1% on MCP-Atlas Tool Orchestration.
| Specification | Value |
|---|---|
| Model Name | Claude Opus 4.7 |
| Release Date | April 16, 2026 |
| Context Window | 1,000,000 tokens |
| Maximum Output | 128,000 tokens |
| Input Pricing (per 1M tokens) | $5.50 |
| Output Pricing (per 1M tokens) | $27.50 |
Claude 4.7 reaches 69.4% on Terminal Bench 2.0. It also reaches 54.5% on Terminal Bench Hard. The model supports vision input, tool calling, and extended reasoning. It also supports prompt caching and structured outputs. Prompt caching can cut effective input cost by up to 90% on repeated context. Claude 4.7 is available via Anthropic’s API and AWS Bedrock.
Anthropic reliability and user feedback issues
Anthropic faced significant stability issues in 2026. On June 3, the Claude Status page recorded 8 incidents across multiple models. Users on Reddit criticized Opus 4.7 as a regression. One user called the model "Gaslightus-4.7" because it denied its own errors. In March, a bug caused Claude to forget prior turns in conversations. On April 21, Anthropic removed Claude Code from the Pro plan. This change led to a 5x price jump for users wanting the feature through the Max plan.
Anthropic also faced issues with usage limits. The limits are stricter than ChatGPT’s limits, especially on free plans. In early 2026, Anthropic’s services ran into compute constraints. This happened after a massive influx of users. Claude 4.6 was the first Opus-class model with a 1M token context in beta. However, the model also had bugs that caused it to forget prior turns. Anthropic released Opus 4.8 on May 28 to address the blind spots of Opus 4.7. Can Anthropic regain user trust through these rapid releases?
The 2026 frontier with Opus 5 and GPT-5.6
The 2026 AI landscape is task-fragmented. Different models lead in different categories. Claude Opus 5 leads in human-preference frontend coding. GPT-5.6 Sol leads in agentic terminal work. Claude Opus 5 reached 42/42 on the 2026 International Mathematical Olympiad problems. It also reached 97.0% on SWE-bench Verified. GPT-5.6 Sol reached 96.2% on the same benchmark. Kimi K3 reached 93.4% on SWE-bench Verified. Claude Opus 5 also quadrupled the ARC-AGI-3 record.
Claude Opus 5 is a specialized model. It is different from the earlier Opus models. Claude Fable 5 leads in repository-level coding. Claude Opus 5 is the strongest in certain human-preference tasks. The frontier includes models like Kimi K3 and various GPT versions. The choice of model depends on the specific task.
Context, pricing, and scaling requirements
GPT-5 has a context window of 400,000 tokens. It has an output window of 128,000 tokens. Claude 4.7 has a context window of 1,000,000 tokens. It has a maximum output of 128,000 tokens. Claude 4.7 costs $5.50 per million input tokens via Requesty. It costs $27.50 per million output tokens via Requesty. You might find the cost disparity between these two providers significant when scaling production workflows.
Claude 4 offers a hybrid workflow. It provides Claude Desktop for developers and enterprises. GPT-5 is available through the OpenAI API and Azure OpenAI Service. Pricing for GPT-5 is token-based. Usage at scale can add up quickly. Both providers offer multi-cloud options. OpenAI works with Azure. Anthropic works with AWS and Google Cloud. This flexibility helps enterprises meet regional compliance requirements.
Coding and agentic benchmark comparisons
Coding performance varies between the models. Claude Opus 4.5 reaches 76.8% on SWE-bench Verified. Claude Sonnet 4 and Opus 4 reach between 72% and 80% accuracy on SWE-bench. GPT-5 reaches 74.9% accuracy on SWE-bench Verified. GPT-5.6 Sol reaches 96.2% on the same benchmark. Claude Opus 5 reaches 97.0% on SWE-bench Verified.
GPT-5 also performs well on multi-language code editing. It achieves 88% accuracy on the Aider Polyglot benchmark. Claude 4 focuses on verbal reasoning. Claude 4 is competitive on everyday reasoning. It trails GPT-5 in high-precision domains like math Olympiad problems. Claude 4 achieves 78% on AIME and over 80% on GPQA. GPT-5 edges ahead in precision.
Model fragmentation and the routing verdict
The industry is moving toward task-specific models. GPT-5 is not the single dominant leader for all tasks. Claude Opus 5 leads in coding, but GPT-5.6 Sol leads in terminal work. A two-model router can recover the full per-benchmark gain of a six-model oracle. The frontier is no longer a single model choice.
Users often find that different models serve different needs. GPT-5 is better for complex reasoning and multimodal tasks. Claude 4 is better for long-context tasks and verbal reasoning. Large organizations manage both models through AI gateways. These gateways handle provisioning, routing, and governing. They also provide unified cost and latency tracking. The most effective way to use these models is to treat them as complementary assets.
