Claude Opus 4.7 Literal Mode: Why Your Old Prompts Suddenly Break

What You'll Learn

• What Anthropic actually documents about Opus 4.7’s more literal behavior
• Which legacy API settings can produce 400 errors after migration
• How tokenizer, thinking display, effort, and output-length changes affect applications
• How to test and repair prompts without inventing a universal failure pattern

Claude Opus 4.7 literal mode is a convenient phrase for a real but narrower migration issue. Anthropic says Opus 4.7 interprets prompts more literally and explicitly than Opus 4.6, especially at lower effort levels. It says the model does not silently generalize an instruction from one item to another and does not infer requests the user did not make.

That does not mean Claude has stopped understanding language. It means old prompts can expose assumptions that were never written down. A request such as “clean up these examples” may no longer imply the same scope, formatting, or exception handling that an earlier model happened to infer. A workflow can feel broken even when the new behavior is closer to the literal text.

There is a separate API migration layer. Anthropic documents that non-default sampling parameters, manual thinking budgets, and assistant prefills can produce 400 errors on Opus 4.7 or later. It also documents a new tokenizer, an omitted thinking display by default, changes in effort calibration, and different tool-use behavior.

The practical fix is not to blame every unexpected answer on a model downgrade or to add a large amount of defensive prose. First identify whether the failure is an API rejection, a prompt-scope change, an output-format change, a token-budget issue, or an agent-harness issue. Then apply the smallest verified repair and run a representative test set.

What Anthropic Actually Changed

Anthropic released Claude Opus 4.7 on April 16, 2026. Its official announcement describes Opus 4.7 as a direct upgrade to Opus 4.6 and says the model is substantially better at following instructions. Anthropic warns that prompts written for earlier models can now produce unexpected results because Opus 4.7 takes instructions literally where earlier models interpreted them loosely or skipped parts.

That statement supports a behavioral change, not a secret toggle. Anthropic’s Messages API documentation provides the request-level context for migration. The public documentation does not call the feature “Literal Mode,” and it does not claim that every prompt written for Opus 4.6 will fail. The most accurate framing is that a model upgrade can change how implicit scope and unstated intent are resolved.

Anthropic’s migration guide also describes API-level changes. If an application uses settings that Opus 4.7 rejects, the request can fail before the model produces an answer. If the request is accepted but the output changes, the team should inspect prompt scope, thinking configuration, effort, token limits, and downstream parsing separately.

For another article focused on the same instruction-following behavior, read our Claude prompt migration analysis. This guide concentrates on the operational checks that should happen before a team calls the change a regression.

What “Literal Mode” Means

“Literal mode” is a useful headline because it describes how a prompt can feel after the upgrade. It is not a documented user-facing switch. Anthropic’s migration guide says Opus 4.7 is more literal and explicit than Opus 4.6, particularly at lower effort levels. It says this can improve precision and reduce unwanted extra work in carefully tuned pipelines.

Consider the difference between “rewrite the examples” and “rewrite every example in the supplied list.” The first sentence leaves the target set unstated. A model may choose the example nearest to the instruction or ask for clarification. The second sentence defines the scope. The prompt has become more precise, not more verbose for its own sake.

The same issue appears with “use the usual format.” A person who knows the application may know that the answer must be JSON with three keys. The model sees no such contract unless the prompt or schema supplies it. When the downstream parser is strict, the format is part of the task, not an optional style preference.

Developers should therefore distinguish inference from invention. A model can use context to understand a request while still refusing to invent an unspoken scope, field, or approval. The more consequential the action, the more valuable an explicit contract becomes.

Which API Requests Can Fail

Anthropic’s migration guide gives concrete API checks for Opus 4.7 and later models. Setting temperature, top_p, or top_k to a non-default value returns a 400 error. The safest migration path is to omit those parameters from the request payload. The Python SDK may also reject these fields at the type level depending on its version.

Manual extended thinking with thinking: {"type": "enabled", "budget_tokens": N} is also no longer supported on Opus 4.7 or later and returns a 400 error. The documented replacement is adaptive thinking with the effort parameter, or no thinking field when the application is intentionally using the model’s ordinary request behavior.

Assistant-message prefilling returns a 400 error as well. Anthropic recommends system prompt instructions, structured outputs, or output configuration instead. A migration should search the actual request builder, not only the prompt text. A parameter hidden inside a shared helper can affect many apparently unrelated features.

These are conditional failures. An application that never sends non-default sampling parameters, manual thinking budgets, or assistant prefills will not break for those reasons. Avoid publishing a universal count of broken integrations when the public documentation only identifies affected configurations.

Manual Thinking and Adaptive Effort

Opus 4.7 and Opus 4.6 behave differently from later models when the request omits a thinking field. Anthropic’s migration guide says requests without a thinking field run without thinking on Opus 4.7. Later models can have adaptive thinking on by default, so a future model migration needs a separate audit rather than an assumption that all Claude versions share one setting.

For Opus 4.7, manual thinking budgets are not a supported replacement for ordinary prompt design. If an application previously used a manual budget, move to the documented configuration for the target model and test the output under the selected effort level. Do not copy a later-model example into an Opus 4.7 integration without checking the model-specific contract.

Effort is a control over how much work the model allocates to a task. The migration guide says Opus 4.7 respects effort levels strictly, especially at the low end. Low or medium effort can help latency and cost, but a moderately complex task may be under-thought. Raising effort is a more direct fix than adding vague instructions such as “try harder.”

Keep the effort choice in the evaluation record. A prompt that works at high effort may not work at low effort. If the product exposes effort to users, test the lowest supported setting that the workflow can tolerate and define when the application should use a higher setting.

Tokenizer and Token Budgets

Anthropic says Opus 4.7 introduced a new tokenizer. The same text may use roughly 1x to 1.35x as many tokens compared with models before Opus 4.7, depending on content and workload shape. This is not a promise of a fixed 35 percent increase for every request. Long prompts, code, structured data, tool results, and language mix can change the result.

Re-baseline token counts with the actual endpoint and workload. Do not rely on a fixed character-to-token ratio or a test prompt that excludes the tool results used in production. Record input tokens, output tokens, number of turns, tool calls, latency, and retries.

The output limit is a hard boundary. If thinking, tool calls, and visible text share the available budget, a setting that was comfortable before the tokenizer change may cut off a response. Add headroom based on measured traffic, then protect downstream systems with structured validation and a clear retry or review path.

Anthropic’s token-counting documentation is the right place to check endpoint behavior. Our AI workflow analysis covers the related question of how tool loops change the cost of a model task.

Prefill and Thinking Display

Assistant prefilling is often used to force a response to begin with a JSON object, a label, or a particular section. Anthropic documents that assistant-message prefilling returns a 400 error on Opus 4.7 and later models. Replace that pattern with a system instruction, structured output, or output configuration supported by the target model.

Thinking display is a different issue. Anthropic says thinking blocks still appear in the response stream on Opus 4.7 and later, but the thinking field is empty unless the application explicitly opts in. The default is omitted. A user interface that expects visible progress may therefore show a long pause before the final answer.

If the product needs summarized thinking progress, set the documented display option and make sure the UI handles the response shape. Do not expose raw internal reasoning as a product promise. A summarized status stream and a final answer are sufficient for many interfaces.

Test streaming and non-streaming paths separately. A migration can pass a simple synchronous test while a streaming client fails because it assumes a particular thinking block, stop reason, or event order.

Scope and Instruction Following

Most “literal mode” complaints are prompt-contract problems. Make the target set explicit. Say whether a rule applies to one item, every item in a list, a named file, a directory, or the entire conversation. State exclusions and what should happen when an item does not match.

Define the output before describing the background. If the consumer expects one JSON object per item, name the keys, value types, ordering, and error value. If the consumer expects prose, state the paragraph count, tone, and whether examples are required. A small example can remove ambiguity that a general instruction leaves behind.

Write down assumptions that change the result. Include the time zone for dates, units for measurements, source priority for conflicts, and the policy for missing fields. If the model should ask a question rather than infer an answer, say so.

Prompt changes should be tested against desired behavior, not against the exact output of the previous model. Anthropic’s prompt-engineering guidance is a useful primary reference for structuring those instructions. An older response may contain accidental helpfulness that the product never formally required. Preserve the behavior that matters and document the rest.

For the surrounding agent ecosystem, see our coding-agent comparison and MCP security checklist. Model capability and tool permission should be evaluated as separate controls.

Response Length, Tone, and Progress

Anthropic says Opus 4.7 calibrates response length to how complex it judges the task to be. Simple lookups may be shorter, while open-ended analysis may be longer. If a product relies on a particular length or voice, the prompt should state that requirement and provide a positive example.

The migration guide also describes Opus 4.7 as more direct and opinionated than Opus 4.6. A change in tone is not evidence that the model is broken. It is a behavior difference that should be tested against the product’s editorial standard.

Long agentic traces can also look different. Anthropic says Opus 4.7 provides more regular progress updates and tends to use tools less often by default while reasoning more. If a harness needs a tool at a specific point, tell the model when the tool is required and validate the result after the call.

Do not use prompt instructions that fight the model’s configuration. If concise output matters, request a target length and define what to omit. If progress updates matter, show an example of the desired status style. If a tool call is mandatory, make the condition observable.

Tool Use and Agentic Workflows

Agentic systems magnify small interpretation differences. A prompt that leaves the repository scope open can lead to the wrong files being inspected. A prompt that does not define the stopping condition can produce extra tool calls. A prompt that says “verify everything” without a test command can create unbounded review work.

Define allowed tools, protected paths, approval points, and completion conditions. Keep the first migration run read-only where possible. Require a file diff, test result, or structured report before the agent claims success.

Anthropic says Opus 4.7 tends to spawn fewer subagents by default than Opus 4.6 and uses tools less often while reasoning more. That can be good for a focused task and undesirable for a workflow that expects search or code execution. Prompt the tool policy explicitly and measure whether the resulting action pattern is useful.

Our AI coding agents guide is useful context, but a product should not assume that one model’s default tool behavior transfers to another. Permissions, harness logic, and acceptance tests remain part of the system.

A Safe Migration Checklist

Start with the request builder. Search for non-default temperature, top_p, or top_k values, manual thinking budgets, assistant prefills, and code that assumes visible thinking text. Remove or replace only the settings that the target model rejects. Keep a copy of the old request so the change is reviewable.

Next, build a fixed test set. Include a normal prompt, a boundary case, a missing field, a repeated list, a conflicting instruction, a long input, a tool error, and a parser failure. Run the same set before and after the model change. Compare the output with the desired result rather than with a screenshot of the old response.

Measure the workflow. Record token counts, latency, retries, tool calls, truncation, human corrections, and acceptance-test results. If the application uses streaming, test event handling and thinking display. If the application uses image inputs, track the image path separately.

Finally, deploy incrementally. Keep a rollback model or configuration, route a small share of traffic through the new version, and inspect failures. A migration is complete when the documented API contract, prompt behavior, and downstream validation all pass for the workload that matters.

Prompt Examples That Expose Assumptions

A vague transformation prompt says: “Clean up this customer email and make it better.” A safer version says: “Correct grammar and spelling only. Preserve the customer’s meaning, names, dates, prices, and paragraph order. Do not add facts. Return the revised email and nothing else.” The second prompt defines scope and exclusions.

A list prompt can be made equally clear: “For every item in the supplied array, return one object with `title`, `status`, and `reason`. If `title` is missing, use `status: unknown` and explain the missing field. Do not omit an item.” This tells the model what “every” means and gives the downstream system a stable error path.

An agent prompt should identify the workspace: “Inspect only files under the named directory. Do not edit dependencies. Run the listed test command after the change. Stop and report if the test command is missing.” These are not magic words. They are a contract that a reviewer can test.

Use examples for difficult transformations, but do not rely on examples to override a contradictory instruction. Keep the priority rule explicit. If the prompt grows large, remove repeated prose and keep the decision rules visible.

For a wider model-selection context, see our AI model comparison guide. A comparison is useful only when the evaluated prompts and acceptance tests are comparable.

What Developers Should Take Away

Claude Opus 4.7 did not acquire a documented switch called Literal Mode. Anthropic documented a stronger tendency to follow instructions literally and explicitly than Opus 4.6. That change can expose prompts that depended on unstated scope, invisible formatting rules, or a model’s previous willingness to fill gaps.

Several API changes are concrete. Non-default sampling parameters, manual thinking budgets, and assistant prefills can return 400 errors. The tokenizer may use roughly 1x to 1.35x as many tokens for the same text. Thinking display is omitted by default unless the application opts in. Effort, response length, tone, progress updates, and tool behavior should be evaluated on the real workload.

The correct migration strategy is evidence-based. Inspect the request builder, define the prompt contract, create a fixed test set, measure token and latency changes, validate structured outputs, and deploy incrementally. Do not claim that thousands of prompts broke or that every developer needs the same fix unless a source provides that evidence.

A model that follows the written instruction more closely can be easier to operate when the written instruction is precise. The upgrade is an invitation to make the contract visible, not a reason to add unsupported folklore to the prompt.