Claude vs GPT-4

The title says GPT-4 because that is what people still search. The real 2026 matchup is Claude's Opus 5 family against OpenAI's GPT-5.6 line.

At a glance

Pricing

Anthropic's current API prices per million tokens: Claude Opus 5 at $5 input and $25 output, Claude Sonnet 5 at $2 and $10, Claude Haiku 4.5 at $1 and $5, and the top-end Claude Fable 5 at $10 and $50. Sonnet 5 launched at $2/$10 as introductory pricing with a scheduled increase to $3/$15 in September 2026; Anthropic has since made the lower price permanent. Batch processing halves all of these, and prompt cache hits bill at 10% of the input rate. On the consumer side, Claude Pro is $17 per month billed annually ($20 monthly) and Max starts at $100 per month.

OpenAI's GPT-5.6 line, released in July 2026, prices Sol at $4 input and $20 output per million tokens, Terra at $2 and $12, and Luna at $0.20 and $1.20, with cached input on Sol at $0.40. OpenAI cut Sol's price in late August 2026, so treat these numbers as current but movable. Older GPT-5.x and o-series models remain available, including GPT-5.5 Pro at $30 and $180 for the hardest problems. All figures come from the two vendors' official pricing pages, checked in August 2026.

When to pick Claude

I pick Claude for agentic coding work, and I should disclose that this site itself is maintained with Claude Code, so I am not neutral. The reason I keep choosing it is the harness: Claude Code, the Agent SDK, and MCP make it straightforward to give a model tools, memory, and long-running tasks without building the scaffolding myself. Opus 5 holds up on long multi-step jobs, and Sonnet 5 at $2/$10 is the price-performance point I route most production work to.

Claude also remains my default for writing that has to sound like a person. That is a subjective call, but it has been a consistent one across my client work, and it is exactly the kind of thing you should test on your own material rather than take from a comparison page.

When to pick GPT

Pick OpenAI when distribution and ecosystem breadth matter more than any single benchmark. ChatGPT is where your clients and employees already are, the surrounding tooling reaches non-technical users, and the OpenAI SDK is the default in more codebases than any other.

The pricing spread is also wider. Luna at $0.20 input and $1.20 output undercuts Haiku 4.5 by a wide margin for high-volume simple tasks like classification and extraction, and at the other end GPT-5.5 Pro exists for problems where you will pay $180 per million output tokens to get the best answer available. If your workload lives at either extreme, OpenAI currently prices it better.

What most comparisons miss

Most pages with this title compare models that no longer matter. GPT-4 versus Claude 3 benchmarks still rank in search results years after both were superseded. Check the date on any comparison before trusting a number, this page included: my figures were verified in August 2026 and will drift.

Headline per-token prices mislead too. Caching, batch discounts, and routing between tiers change effective cost by multiples. A team that caches a large system prompt on Claude pays closer to $0.20 than $2 per million input tokens on Sonnet 5. Bill projections need your actual traffic shape, not the pricing table.

Finally, the two lineups have converged: same tiering, similar context windows, similar tool calling. The durable differences are the agent ecosystems, not raw scores. Swapping the model string in an API call is cheap, while swapping an agent harness is not, and that is where the real lock-in lives.