Claude Sonnet 5 API: Online text generation, tutorials and integration guide

Project coding: implement features and make focused changes around a clear plan

We believe Claude Sonnet 5 can implement features, perform focused refactoring, or fix issues once the plan is clear. One customer told us that he generally used Sonnet for coordination and implementation, with Opus handling architecture and difficult problems. The project was ultimately completed, although he had to switch to Opus twice to keep going.

We think this division of work is worth considering: clarify interfaces, the scope of impact, and acceptance criteria before asking Sonnet 5 to begin. If an architectural change is needed, or the tests still fail after several rounds of changes, we recommend revisiting the plan and considering a different model. A lower unit price does not necessarily mean lower costs after repeated retries. Of course, this is only one customer's project, and we still recommend testing your actual business scenarios before switching to Sonnet 5.

Tool collaboration: recover from error feedback and continue execution

One of our customers shared his experience integrating an agent. According to him, Sonnet 5 could handle complex editing requirements in a single pass. After encountering an unclear error, it would call tools again to inspect the data structure, look for input problems, and attempt recovery. However, he also mentioned that the model once skipped another input it could not process without telling the user.

We believe that when using Sonnet 5, you need to check both the model's response and whether the data ultimately written is correct. Being fast does not mean the response meets your requirements.

Long-document analysis: work across specifications, code, and business materials

We use Sonnet 5's long context to cross-reference multiple sources: comparing versions of requirements, checking whether code follows interface agreements, or finding business conditions scattered across documents. In Artificial Analysis's AA-LCR v1.1 comparison, Sonnet 5 Max scores approximately 82%. This evaluation uses documents of roughly 10,000 to 100,000 tokens to assess whether models can extract information across passages, reason, and synthesize conclusions. Sonnet 5 long-document results · AA-LCR methodology

We recommend defining the question first, then organizing the input materials, and retaining file names, sections, and key quotations in the answer so that you can check whether it meets your requirements. When the input approaches a million tokens, we recommend separately checking for omissions. The 82% score comes from the document-length range described above. In actual business scenarios, however, we do not recommend inputs approaching a million tokens: in our own experience, this clearly affects output quality.

Business writing: turn established facts into explanations your readers can understand

One customer shared his use case with us: he uses Sonnet 5 in an API automation workflow to write customer emails and text messages, and to turn Git diffs into change summaries that nontechnical readers can understand. He rated these uses positively, particularly noting that the writing was readable and felt human, and that the model followed the system prompt's requirements.

We believe different business scenarios call for different standards of readability. Product updates, support ticket replies, and help documents tend to be less demanding in this respect. General articles or personal reflections, on the other hand, may require people to do most of the editing and refinement.

Interface development: use screenshots and design specifications to generate editable frontends

We would prioritize Sonnet 5 for modifying existing interfaces or building pages with clearly defined design requirements. Anthropic's prompting guide notes that overly open-ended frontend requests may lead the model to repeatedly use similar default visual styles. To get a different result, provide concrete design constraints that limit the model's improvisation. Official frontend prompting guidance

Feedback in our customer group is also mixed: some said that when they used Sonnet 5 to rebuild interfaces previously created with Sonnet 4.6, the visual results did not meet their expectations. Other customers, however, said Sonnet 5 performed quite well.

We therefore recommend providing page screenshots, component specifications, font requirements, and layout requirements together. Once the page is ready, check functionality, fidelity to the design, and how it displays on different screens separately. If the project already has a mature design system, we recommend reusing existing components and specifying the changes clearly rather than leaving the model to design and improvise from scratch.

Terminal tasks: approximately 14% in the independent Terminal-Bench 4.0 evaluation

In Artificial Analysis's like-for-like comparison, Sonnet 5 Max scores approximately 14% on Terminal-Bench 4.0, compared with approximately 3% for Sonnet 4.6 Max. The evaluation includes 66 terminal tasks, uses mini-swe-agent, repeats each task three times, and calculates the average single-attempt pass rate. A task counts as successful only when all its tests pass. Model score comparison · Terminal-Bench 4.0 methodology

We believe Sonnet 5 improves on the previous generation in this evaluation, but still cannot complete many difficult terminal tasks. We do not recommend Sonnet 5 for particularly complex tasks. When a task is clearly defined and broken into small units, however, Sonnet 5 may be a good choice.

Project coding: CursorBench 4.0 gains 3.3 percentage points from High to Max

CursorBench 4.0 uses project tasks drawn from real development workflows. Some of the organizer's published Sonnet 5 results are shown below, with costs calculated from the corresponding tasks' token usage and public prices. Original CursorBench 4.0 leaderboard

We pay closer attention to the cost comparison between high and max: the score rises by 3.3 percentage points, but the average task cost more than doubles. We recommend comparing medium and high first for everyday development. If a problem repeatedly gets stuck in your scenario, we recommend using Opus 5.5 or Opus 5 to solve it rather than raising the thinking level further.

Knowledge work: 1465 Elo on GDPval-AA v2.1

In Artificial Analysis's GDPval-AA v2.1 comparison, Sonnet 5 Max scores 1465 Elo, versus 1234 Elo for Sonnet 4.6 Max. This evaluation uses 220 tasks spanning 44 occupations and 9 industries. Models work in a tool environment to submit deliverables such as documents, spreadsheets, and presentations, which are then scored through anonymous pairwise comparisons. Sonnet 5 and previous-generation results · GDPval-AA v2.1 methodology

We believe this result can help assess whether the model can produce documents, spreadsheets, and other files for other people to use. When evaluating report preparation, business analysis, or office assistants, we recommend giving it a complete task directly: did it use the necessary materials, are the numbers correct, can the files be opened, and can the intended readers use them?

Business automation: approximately 37% goal completion on AutomationBench-AA

In Artificial Analysis's comparison, Sonnet 5 Max scores approximately 37% on AutomationBench-AA, versus approximately 20% for Sonnet 4.6 Max. This evaluation covers 657 tasks performed in simulated business applications such as email, spreadsheets, and CRM systems. Model score comparison

The score first measures the proportion of goals completed in each task, then averages those proportions; a violation of task constraints receives a zero. Therefore, 37% does not mean that "37% of complete business workflows succeeded from start to finish." It also differs from the full-task completion rate on Zapier's own leaderboard, and the two measures must not be mixed. AutomationBench-AA scoring explanation

We recommend Sonnet 5 for business automation steps with clear, simple boundaries that require fast responses. If you need it to solve more complex problems, such as making decisions, we do not recommend Sonnet; you can use Opus 5.5 or Opus 5 instead.

Overall index and generation speed: a score of 38 and approximately 80.5 tokens per second

Artificial Analysis's current model page gives Sonnet 5 Max a score of 38 on Intelligence Index v4.3.2 and measures Anthropic API output speed at approximately 80.5 tokens per second. The weighted average cost per task for the same index is US$5.09. Artificial Analysis model evaluation

For frequently used interactive features, we recommend trying lower thinking levels; comparing waiting time and output quality is essential. If the output at medium/high also fits your business scenario, I recommend using Sonnet 5.

Compared with Claude Sonnet 4.6: stronger capabilities, but higher task costs at Max

In Artificial Analysis's Max-configuration comparison, Sonnet 5 and Sonnet 4.6 score 38 and 30 on the overall index, with weighted average costs per task of US$5.09 and US$2.49, respectively. Their public input/output prices are US$2/US$10 and US$3/US$15 per million tokens, respectively. Sonnet 5 has lower per-token prices, but higher task costs in this evaluation. Sonnet 5 versus Sonnet 4.6

For teams still using Sonnet 4.6, we recommend first checking whether the upgrade improves tool collaboration and project execution, while readjusting thinking levels and output budgets. If Sonnet 5 can reliably complete tasks at medium or high, we do not think there is a need to keep using the highest level.

Compared with GPT-6 Astra: reevaluate complex engineering tasks using actual consumption

In Artificial Analysis's Max-configuration comparison, GPT-6 Astra and Sonnet 5 score 53 and 38 on the overall index, with Terminal-Bench 4.0 scores of 59% and 14% after rounding as displayed on the page. Although Astra's per-token prices are higher, its weighted average cost per task in this overall evaluation is US$3.26, below Sonnet 5's US$5.09; its corresponding average output per task is also lower. GPT-6 Astra versus Sonnet 5

If your work mainly involves complex code changes, terminal operations, or engineering tasks requiring repeated verification, we recommend a close comparison with GPT-6 Astra. In our experience so far, GPT-6 Astra is a better choice than Sonnet 5 for complex engineering.

Compared with DeepSeek V4 Pro: a useful cost comparison for batch text tasks

In Artificial Analysis's Max-configuration comparison, DeepSeek V4 Pro 0813 and Sonnet 5 score 36 and 38 on the overall index, and approximately 80% and 82% on AA-LCR v1.1 long-document tasks; their weighted average costs per task are US$0.67 and US$5.09, respectively. Their input capabilities also differ: the DeepSeek V4 Pro evaluated here accepts text input, while Sonnet 5 supports text and image inputs. DeepSeek V4 Pro versus Sonnet 5

For organizing information, extracting text, and drafting in batches, we recommend including DeepSeek V4 Pro in your cost tests. If you want to test the multimodal capabilities of DeepSeek V4 PRO and Sonnet 5, we strongly recommend using real scenarios from your business. In customer feedback, some users found that Sonnet 5's multimodal capabilities did not meet their business requirements, while others said that DeepSeek V4 PRO was not capable enough to handle their work.

How do I call the Claude Sonnet 5 API through RequestBeam?

We recommend first signing in to RequestBeam and creating an API Key, then reading our documentation or asking AI to help you configure the connection.

We recommend completing a minimal generation test through the online tool first, then integrating according to RequestBeam's current Claude Sonnet 5 API documentation. This lets you verify whether the model meets your business needs before integration and helps avoid mistakenly carrying over endpoint paths and parameters from other providers into RequestBeam.

Why do old thinking or sampling parameters return 400 after migrating from Sonnet 4.6?

We recommend first checking thinking, budget_tokens, temperature, top_p, and top_k in your request. Sonnet 5 no longer supports the legacy thinking: {"type":"enabled","budget_tokens":N} format, and non-default sampling parameters also return 400. When migrating, we recommend removing these old settings and using output_config.effort to control reasoning effort instead. Sonnet 5 migration and parameter documentation

When thinking is omitted, Sonnet 5 enables adaptive thinking. If you really need to turn it off, set thinking: {"type":"disabled"}. We recommend explicitly setting effort according to your business needs, and controlling tone, formatting, and content scope through prompts and reference examples. Official thinking configuration

Why does a successful request have no answer text while output tokens still increase?

We recommend checking stop_reason and the returned content blocks first. Sonnet 5 does not display thinking text by default, so the thinking field may be empty even though thinking tokens still count toward output usage. When reading the answer, we recommend extracting text according to each content block's type rather than always reading content[0].text. Thinking display and billing explanation

Also check the output budget. max_tokens limits the combined output of thinking and answer text. With a high effort and a small budget, the model may reach the limit before finishing the answer. When you see stop_reason: "max_tokens", we recommend increasing the output budget or trying a lower thinking level, then checking whether the answer is complete. Sonnet 5 output budget documentation

Why does the task not continue after the model returns a tool call?

When using client-side tools, a model response with stop_reason: "tool_use" means it is waiting for the application to execute the tool and return the result. The application must first add the original assistant content to the message history, then return the execution result with tool_result. The tool_use_id must match the earlier call before you can request the model again. Official tool-use workflow

We recommend preserving all returned content blocks, particularly the relevant blocks and signatures when thinking is enabled, rather than retaining only the text visible to users. For integration testing, we recommend first completing one cycle of "request a tool, execute the operation, return the result, generate an answer" before expanding to multi-tool workflows. If tool execution fails, return the error information as well so that the model can adjust its next step. Thinking content block preservation requirements

Usage-based billing and model cost optimization

Reliable service at prices lower than official rates. Top up on demand and use what you pay for.

Unified AI models and data APIs, with multi-model access in one place

One API key connects multiple AI models and data services, reducing integration costs and letting you switch with one click. Business applications need both AI and data.

Call logs and application access controls, with clarity on every call

Every API call is recorded in your dashboard. Create separate API keys for different applications.

Model capabilities without compromise

We do not water down our models. We cannot offer you the cheapest, but we are committed to offering you the best.

Prepare your account and test online

Sign in, confirm your balance, and create an API key.

Try it in the Playground

Explore the model's capabilities in the Playground to find the best way to use it.

Integrate Claude Sonnet 5 into your business

Follow our developer documentation to integrate Claude Sonnet 5 into your business.

Future model upgrades

When a new model is released, simply replace the model name with the new one to upgrade.