Claude Opus 5 (Adaptive Reasoning, Medium Effort) vs Gemini 1.5 Pro (Sep '24): The Ultimate Performance & Pricing Comparison

Claude Opus 5 is the safer choice for a new developer project because it is active, documented, and materially stronger on the supplied coding evaluation. Anthropic identifies claude-opus-5 as the API model ID and stable alias, while Medium Effort is configured with effort: "medium", not exposed as a separate API model. The official model overview also documents access through the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Gemini 1.5 Pro (Sep '24) is harder to select responsibly because the current Gemini API model documentation no longer presents an active model card for that specific version. The supplied evidence does not confirm a current endpoint, version-specific output limit, or active commercial path. That difference matters more than historical familiarity: a model that cannot be confidently provisioned is a deployment risk before quality is even tested.

Data provided by Artificial Analysis.

Claude Opus 5 offers the stronger overall selection case because its documented operating model aligns with the evaluation data and with production requirements. Its Artificial Analysis Intelligence Index is 56.3, compared with 10 for Gemini 1.5 Pro, while its Coding Index is 74.3 compared with 23.6. These figures do not prove that every prompt will favor Claude, but they establish a large measured gap for the developer-oriented decision described here.

The main unresolved question is not whether Gemini 1.5 Pro once had useful long-context capabilities. It is whether a developer can still call this exact version with documented behavior and a current price. The supplied sources do not answer that question positively. Google's current pricing page does not list an active Gemini 1.5 Pro price, so historical comparisons should not be treated as procurement guidance.

Claude Opus 5 is the better performance bet for repository-scale coding because its measured coding score is 74.3 and its median output speed is 54.838 tokens per second. The chart below provides the numerical comparison, so the practical issue is how those values affect workflow design.

A stronger coding score should matter most when the task requires several dependent actions: inspecting an unfamiliar repository, changing multiple files, running tests, interpreting failures, and revising the implementation. Anthropic explicitly positions Claude Opus 5 for complex agentic coding, long-horizon work, code review, multi-file development, and multi-agent collaboration in its Opus 5 update. Those claims fit the kind of work where a model must preserve constraints across many tool calls.

The speed result needs a narrower interpretation. Gemini 1.5 Pro has no supplied median output-speed value, so the comparison cannot establish that Claude is faster in every interactive setting. Both models show 0.3 seconds for the supplied latency metric, which suggests that first-response responsiveness does not separate them in this snapshot. Streaming speed and end-to-end task completion are different variables. A model that reasons, tests, and revises more effectively may finish a larger task with fewer corrective turns, but the supplied data does not measure that directly.

Claude also has documented output and reasoning interactions that developers must control. Thinking tokens share the max_tokens ceiling, and default reasoning can reduce visible output if the ceiling is not adjusted. The official documentation warns that disabling thinking can occasionally expose tool-call text or internal XML-like tags. These are integration concerns, not proof of poor capability, but they make a small evaluation harness essential.

Gemini's missing current version documentation prevents an equivalent failure-mode analysis. The supplied sources do not provide reproducible Gemini coding tests, current limits, or community evidence for this exact release. A team considering Gemini should therefore treat its performance result as historical or environment-dependent until it can verify an endpoint and run the same tasks.

Claude Opus 5 is cheaper on every supplied price measure, and its documented billing makes the cost decision easier to operationalize. The chart below shows the prices, so the important question is when token price translates into total application cost.

Claude costs $10 per 1M blended tokens, compared with $15 for Gemini 1.5 Pro. Its input price is $5 per 1M tokens and its output price is $25, compared with Gemini's $10 input and $30 output prices. Claude therefore has the lower listed rate for both prompt-heavy and response-heavy workloads. Anthropic also documents prompt caching at $0.50 for cache hits, with separate write prices of $6.25 for five-minute storage and $10 for one-hour storage on the official pricing page. Repeated repository context may benefit from that structure, although actual savings depend on cache eligibility and request patterns.

A lower token rate does not automatically mean a lower monthly bill. If Claude's adaptive reasoning produces longer responses, extra validation, or more delegated subtasks, an application that pays for every token could spend more than a simple blended comparison suggests. Anthropic explicitly warns that default responses and delivery documents may be longer, with more progress narration, self-verification, and delegation. Developers should cap unnecessary output, choose the intended effort level, and measure completed task cost rather than response cost alone.

Gemini 1.5 Pro creates a different budget risk. Google's current pricing documentation does not list a current price for this version, so the $15 benchmark value cannot be assumed to be a live purchasable rate. The supplied evidence also does not confirm whether an old configuration would route successfully, fail, or require migration. For a new product, an undocumented price is not a discount. It is an unpriced dependency that can invalidate a forecast.

Claude Opus 5 should be the default pick for a new coding product, while Gemini 1.5 Pro is reasonable only when an existing system can already verify its endpoint and behavior. The recommendation follows from the combination of active documentation, lower supplied prices, and stronger measured coding and intelligence scores.

Choose Claude Opus 5 for an agent that edits repositories, performs code review, handles multi-file changes, or works through a task over many iterations. Anthropic's release announcement describes the model as aimed at complex agentic coding and enterprise work. The supplied evaluation gives it a Coding Index of 74.3, which supports that positioning more directly than the historical Gemini material supports a current version choice.

Choose Claude with explicit controls. Use the real API ID, claude-opus-5, and set effort to medium when reproducing the compared configuration. Do not invent a separate claude-opus-5-medium endpoint. Decide whether thinking should remain enabled, size max_tokens for both reasoning and visible output, and add tool-call validation. These steps address documented behavior rather than speculative tuning.

Consider Gemini 1.5 Pro only for a constrained migration, a legacy integration, or a validated internal experiment. Its historical long-context reputation may still be relevant to a team with an existing Google deployment, but the supplied model directory does not confirm an active model card or current endpoint for this exact version. The research brief also found no reproducible community coding tests for the release.

The evidence is insufficient to declare a universal winner for multimodal workloads, audio or video-heavy prompts, or tasks that depend primarily on a very large context window. Claude's official materials document text and image input, while historical Gemini materials describe text, image, video, and audio input. The current Gemini source does not preserve enough version-specific detail to make a fair capability comparison. Validate those workloads directly before committing.

Claude Opus 5 is easier to put into production today because its identity, limits, platforms, and pricing are documented in active official materials. The questions below focus on risks that the headline scores cannot resolve.