AI / Claude OPUS5 Interview questions

Claude Opus 5 is Anthropic's current Opus-tier model, positioned as an everyday, general-purpose model for coding, agentic work, and enterprise knowledge tasks, released July 24, 2026.

It's designed to approach the intelligence of Anthropic's top-tier Claude Fable 5 model on many tasks while costing about half as much, making it the more cost-efficient default for most day-to-day use unless a task specifically calls for Fable 5's extra capability.

It's available as claude-opus-5 through the Claude API, and also through Amazon Bedrock, Google Cloud, and Microsoft Foundry, alongside continued availability of the prior Opus 4.8 model.

Claude Opus 5 was released on July 24, 2026, about eight weeks after Claude Opus 4.8, which itself had shipped on May 28, 2026 - a notably tighter release cadence than earlier Opus generations.

Its launch came alongside two other platform updates on the same day: mid-conversation tool changes, letting developers add or remove tools between turns without invalidating the prompt cache, and automatic fallbacks on the API, routing safety-classifier-flagged requests to another model instead of blocking them.

The release followed shortly after Claude Fable 5 and Claude Mythos 5, Anthropic's Mythos-tier models, which had themselves launched on June 9, 2026.

Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, identical to the pricing of its predecessor, Claude Opus 4.8.

A separate Fast mode, offered as a research preview, costs roughly double the standard rate - $10 per million input tokens and $50 per million output tokens - in exchange for output that's about 2.5 times faster.

Because pricing didn't change from Opus 4.8, the practical cost impact of migrating comes less from the sticker price and more from behavioral differences, like thinking now running by default, which can change how many tokens a given request actually consumes.

On the Claude API, Claude Opus 5 is addressed with the model string claude-opus-5, replacing claude-opus-4-8 for teams migrating from the prior version.

On Amazon Bedrock it's available as anthropic.claude-opus-5, and it's also reachable through Bedrock's InvokeModel API on the same underlying infrastructure, though Bedrock's separate legacy ARN-versioned integration doesn't include it.

On Google Cloud and Microsoft Foundry, it's available under the same bare claude-opus-5 identifier, and Opus 4.8 remains available in parallel on all these platforms, so switching isn't forced.

Claude Fable 5 is Anthropic's top-tier, frontier-intelligence model, sitting above the Opus tier and released as part of the same Mythos-tier generation as Claude Mythos 5.

It shares its underlying model with Claude Mythos 5 but includes additional safety measures specific to biology, cybersecurity, and LLM R&D domains, reflecting a more cautious release posture for Anthropic's most capable model.

Claude Opus 5 is explicitly positioned as coming close to Fable 5's intelligence on many everyday tasks at roughly half the price, though Anthropic still recommends Fable 5 for more demanding or longer-running autonomous projects.

The effort parameter controls how much computational work Claude Opus 5 puts into a request - governing thinking depth as well as the depth of tool calls and arguments, not just visible response length - across levels typically described as low, medium, high, xhigh, and max.

Claude Opus 5 defaults to high effort on the API, and the specific amount of work each level represents was recalibrated relative to Opus 4.8, so a setting carried over unchanged from the prior model won't necessarily produce the same cost or behavior.

Because effort mainly governs thinking and tool-call depth rather than visible answer length, lowering it to save cost won't reliably make responses shorter - controlling response length requires an explicit instruction in the prompt instead.

Fast mode is a research-preview option for Claude Opus 5 that trades higher per-token pricing for roughly 2.5 times faster output, aimed at exploration, drafting, and lower-risk loops where speed matters more than squeezing out maximum cost efficiency.

It's currently available only through the Claude API directly, not through Amazon Bedrock, Google Cloud, or Microsoft Foundry, so teams relying on those platforms don't yet have access to it for Opus 5.

It continues a pattern from earlier Opus versions (4.6 through 4.8), which also shipped with their own fast-mode sibling at premium pricing, rather than being a brand-new capability introduced specifically with Opus 5.

Claude Opus 5 has a 1 million token context window, and that 1M figure is both the default and the absolute maximum - there's no smaller context-window variant to opt into instead.

This matches the context window Claude Opus 4.8 and the Mythos-tier models, Fable 5 and Mythos 5, also use by default, keeping context handling consistent across Anthropic's current model lineup rather than varying significantly release to release.

A large, fixed context window like this is particularly relevant for long-horizon agentic tasks, where a session's accumulated tool outputs, code, and conversation history can grow substantially before a task completes.

Claude Opus 5 supports a maximum of 128,000 output tokens per response, the same ceiling used by its predecessor Opus 4.8.

Because thinking now runs by default on Opus 5, and thinking tokens share the same max_tokens budget as the visible reply, a request that doesn't account for this can end up with less room for the actual answer than it would have gotten on Opus 4.8 under the same max_tokens setting.

This makes it worth explicitly checking max_tokens budgeting when migrating from Opus 4.8, rather than assuming the same numeric value will behave identically now that thinking consumes part of that same budget by default.

Thinking, also called reasoning, is Claude working through a problem step by step before producing its final answer, and on Claude Opus 5 it runs by default on every request rather than needing to be explicitly turned on.

This is a first for the Opus model family - every Opus version before Opus 5 had thinking off by default unless a request explicitly set it, such as with thinking: {"type": "adaptive"} - though Claude Sonnet 5 had already introduced this on-by-default behavior before Opus 5 did.

The amount of thinking Claude Opus 5 actually performs on a given turn is adaptive and controlled by the effort parameter, so higher effort levels generally allow deeper reasoning, at a correspondingly higher token cost.

The minimum prompt length eligible for caching on Claude Opus 5 is 512 tokens, down from 1,024 tokens on Claude Opus 4.8.

This means system prompts or other cacheable content that were previously too short to qualify for caching on Opus 4.8 can now create cache entries on Opus 5 automatically, with no code changes required.

For teams running high-volume agentic workloads with relatively short, frequently-reused system prompts, this lowered threshold can translate into a straightforward cost reduction simply by switching to the new model.

Mid-conversation tool changes is a beta capability that lets developers add or remove tools available to Claude between turns of an ongoing conversation, without invalidating the prompt cache the way changing the tool list previously would have.

Before this, an agentic session generally needed to commit to a fixed tool list for the life of the conversation, since changing it broke caching and could force expensive cache misses on subsequent turns.

This is particularly useful for agents that need to progressively unlock or restrict capabilities as a task unfolds - for example, granting access to a deployment tool only once a task reaches a later stage - without paying a caching penalty for that change.

Automatic fallback is a beta feature where API requests that get flagged by Claude Opus 5's or Claude Fable 5's safety classifiers can be automatically routed to another model instead of simply being blocked outright.

With automatic fallback enabled, a request effectively always routes to the best available model that can serve it, rather than the developer's application receiving an outright refusal or error for a flagged request.

This gives developers a way to reduce friction from occasional over-triggering safety classifications, without disabling safety review entirely - the request is redirected rather than the classifier being bypassed.

Claude Opus 5 is available directly through the Claude API, as claude-opus-5, through Amazon Bedrock, as anthropic.claude-opus-5, through Claude on Google Cloud, as claude-opus-5, and through Claude in Microsoft Foundry.

It's also the default model on Claude Max and the strongest model selectable on Claude Pro, so consumer-facing Claude.ai users encounter it without needing to go through the API at all.

Claude Opus 4.8 remains available in parallel across all of these same platforms, so switching to Opus 5 is optional rather than a forced migration for existing integrations.

On Claude Max, Anthropic's premium consumer subscription tier, Claude Opus 5 became the new default model upon release, replacing Claude Opus 4.8 in that role.

On Claude Pro, Claude Opus 5 is the strongest model available to select, positioning it as the top consumer-facing option on that plan specifically.

This reflects Anthropic's framing of Opus 5 as an everyday, general-purpose model well suited to typical subscriber usage, rather than reserving that consumer-facing default role exclusively for the more capable Fable 5 model.

Claude Mythos 5 shares the same underlying model as Claude Fable 5, both belonging to Anthropic's Mythos tier, which sits above the Opus tier in capability.

The two differ mainly in safety posture: Fable 5 includes additional safety measures specific to biology, cybersecurity, and LLM R&D, while Mythos 5 itself is not currently available to the general public and is instead used by a small number of trusted organizations as part of Anthropic's Project Glasswing.

Public benchmark comparisons note that Claude Opus 5 trails Mythos 5 specifically on certain cybersecurity evaluations, even though it approaches Fable 5's general capability on many other tasks.

On Claude Opus 5, thinking is already enabled by default for requests that don't configure it explicitly, a change from every prior Opus model, where thinking needed to be explicitly turned on with something like thinking: {"type": "adaptive"}.

Because of this, developers migrating from Opus 4.8 generally don't need to add anything to get thinking behavior on Opus 5 - if anything, they need to actively opt out if they want it disabled.

To disable thinking instead, a request can set thinking: {"type": "disabled"}, but this is only accepted at effort level high or below - attempting to disable thinking at xhigh or max effort returns an HTTP 400 error, since Anthropic requires thinking to stay on at those higher effort levels.

Claude Opus 5 defaults to high effort on the API, which Anthropic positions as fitting most everyday coding and chat use cases without further tuning.

This same high default is shared with Claude Fable 5, so both models start from a comparable baseline of effort unless a request explicitly dials it up or down.

Because effort levels were recalibrated between Opus 4.8 and Opus 5, teams that previously tuned an explicit effort setting on Opus 4.8 are advised to re-run an effort comparison on their own evaluations against Opus 5, rather than assuming the same label still costs or behaves the same way.

Subagent delegation refers to Claude Opus 5's tendency, during agentic work, to hand off portions of a task to subagents more readily than earlier Opus models did.

This is one of the behavioral shifts Anthropic's own prompting guidance calls out as worth checking when moving a workflow from Opus 4.8 to Opus 5, alongside longer default response length and more frequent narration.

Because this changes how a multi-step agentic task actually gets broken up and executed, teams with agent orchestration logic tuned around Opus 4.8's delegation patterns may need to review and adjust that logic against Opus 5's more willing delegation behavior.

The baseline step is simple: change the model identifier from claude-opus-4-8 to claude-opus-5 in API requests, since most existing requests continue to work without further changes.

Beyond that string swap, it's worth explicitly checking a small number of behavior changes that can affect correctness or cost: thinking now runs by default and shares the max_tokens budget with the visible reply, disabling thinking is restricted to effort high or below, effort levels were recalibrated, and Opus 5 draws from a separate rate-limit bucket rather than the combined Opus 4.x pool.

On the prompting side, it's also worth reviewing prompts for outdated verbosity or verification instructions carried over from Opus 4.8, since Opus 5's different default behavior around response length and self-verification can interact with those older instructions in ways that waste tokens rather than helping.

The most consequential functional difference is that Claude Opus 5 runs with thinking on by default, while Opus 4.8 ran without thinking unless a request explicitly enabled it - this alone changes both typical token cost and answer quality on unconfigured requests.

Opus 5 also lowers the prompt-caching minimum to 512 tokens (from 1,024), uses a separate rate-limit bucket rather than the shared Opus 4.x pool, adds mid-conversation tool changes and automatic fallback as new beta capabilities, and drops support for Priority Tier, a feature some enterprise customers relied on with Opus 4.8.

On raw capability, Anthropic reports Opus 5 achieving markedly better results on agentic and reasoning-heavy benchmarks like Frontier-Bench, at the same $5/$25 per-million-token pricing as Opus 4.8 - so the upgrade is framed primarily as a capability and efficiency gain rather than a price change.

Anthropic positions Claude Opus 5 as approaching Fable 5's frontier intelligence on many everyday tasks while costing roughly half as much, making it the more cost-efficient default choice for typical coding, analysis, and business-automation work.

On specific published benchmarks, the gap narrows further under higher effort settings - for example, at max effort on CursorBench 3.2, Opus 5's score lands within about half a percent of Fable 5's peak, while still costing half as much per task.

Fable 5 remains the recommended choice for the most demanding or longest-running autonomous work, such as multi-day agentic projects, where its extra capability margin is more likely to matter in practice than it does on shorter, everyday tasks.

Because max_tokens caps thinking and the visible reply together as one shared budget, a request that previously produced no thinking at all on Opus 4.8 now consumes part of that same budget on thinking before Opus 5 even starts on the visible answer.

If an application's max_tokens value was tuned assuming the full budget went to the visible reply, as it did by default on Opus 4.8, that same value on Opus 5 can leave less room than expected for the answer - the request still returns successfully with HTTP 200, but with stop_reason: max_tokens and the answer cut off mid-sentence, rather than throwing an obvious error.

This makes it a quiet breaking change: nothing in the API contract technically fails, so it's easy to miss during migration testing unless truncated responses are specifically checked for on requests that previously fit comfortably within budget.

The effort parameter controls how much Claude Opus 5 thinks, and by extension how deep its reasoning, tool calls, and argument construction go - it's fundamentally about internal computational work, not the length of the final visible answer.

Response verbosity, the actual length of what the user sees, is a separate concern that effort doesn't reliably control - lowering effort to save cost will not dependably make the visible response shorter, which surprises teams expecting effort to double as a length dial the way it might seem to on the surface.

To actually control response length, Anthropic's guidance recommends an explicit instruction in the prompt itself, such as asking for focused, concise answers with minimal caveats, rather than relying on the effort parameter to do that job.

Reach for Fable 5 when a task is at the edge of what's currently possible - the most demanding reasoning, or the largest and most ambiguous multi-day autonomous agentic projects - where Opus 5's slightly smaller capability margin relative to Fable 5 is more likely to actually matter to the outcome.

For the large majority of everyday coding, analysis, and business-automation tasks, Opus 5 is the more cost-efficient default, since it approaches Fable 5's intelligence on many of these tasks at roughly half the price, making the extra Fable 5 cost hard to justify unless the task specifically needs that margin.

It's also worth noting that Fable 5 and Mythos 5 always run with adaptive thinking on with no way to disable it, and carry their own distinct safety classifier behavior, so a task with very specific thinking-control requirements is another factor in the choice, not just raw capability.

Claude Opus 5 verifies its own work unprompted more consistently than Opus 4.8 did, meaning it's more likely to catch and correct its own mistakes without an explicit instruction telling it to double-check.

This creates a specific migration trap: verification instructions written for Opus 4.8, like an explicit "double-check your work" line, now compound with Opus 5's own built-in verification behavior, causing over-verification that spends extra tokens without adding proportional value.

Anthropic's own prompting guidance recommends removing global double-check instructions when migrating, and only scoping an explicit verification instruction to a specific step if that particular step genuinely still needs one, rather than applying it as a blanket rule the way it may have been useful on Opus 4.8.

On Claude Opus 5, thinking can be explicitly disabled with thinking: {"type": "disabled"}, but only at effort level high or below - attempting this at xhigh or max effort returns an HTTP 400 error, since those higher effort levels require thinking to stay on.

On Claude Fable 5 and Claude Mythos 5, adaptive thinking is always on with no disabled mode available at any effort level at all - it's simply not a configuration option the way it is, with restrictions, on Opus 5.

This reflects a broader pattern across Anthropic's current lineup: Opus 5 offers a genuine, if now restricted, choice about whether thinking runs, while the Mythos-tier models treat always-on adaptive thinking as a fixed characteristic rather than a togglable setting.

On Claude Opus 4.8, a request that didn't explicitly enable thinking produced no thinking tokens at all, so the full max_tokens budget was effectively available to the visible reply by default.

On Claude Opus 5, thinking runs by default for the same kind of unconfigured request, and thinking tokens draw from that same max_tokens budget before the visible reply gets its share - so the same numeric max_tokens value now has to cover both, rather than just the answer.

The practical effect is that a max_tokens setting that was comfortably generous on Opus 4.8 can become tight on Opus 5 for the exact same request, since part of that budget is now consumed by reasoning the request wasn't previously spending any tokens on.

Claude Opus 5 tends to narrate more during agentic work than earlier models - announcing what it's about to do before taking an action - and its per-message output during agentic sessions tends to run longer as a result.

Left unguided, this can make agent transcripts noisier than teams may want, especially in production settings where users see this narration directly rather than just the final result.

Anthropic's guidance recommends describing the specific narration cadence wanted, rather than simply forbidding narration outright - for example, asking for one sentence before the first tool call, brief updates only when something important changes, and a final message that leads with the outcome, since positive descriptions of desired style tend to work better than blanket instructions.

Claude Opus 5 draws on its own separate rate-limit bucket, distinct from the combined pool that Opus 4.x models, like 4.6, 4.7, and 4.8, share together.

This means shifting traffic from Opus 4.8 over to Opus 5 doesn't free up headroom on the old shared Opus 4.x bucket, and it also doesn't automatically inherit whatever headroom existed there - the two are tracked independently.

Because of this, teams planning a significant traffic shift to Opus 5 are advised to check their account's specific Opus 5 rate limits ahead of time, rather than assuming existing Opus 4.8 capacity planning carries over directly to the new model.

Fast mode suits exploration, rapid drafting, and lower-risk iterative loops where getting a response back quickly matters more than minimizing cost per token, since it runs at roughly 2.5 times the speed of standard mode at about double the price.

It's less suited to workloads where cost efficiency is the primary concern, or where the specific platform being used doesn't yet support it - currently it's a Claude API-only research preview, unavailable through Amazon Bedrock, Google Cloud, or Microsoft Foundry.

Because it's a research preview rather than a fully stable feature, teams should also weigh that its availability, pricing, or exact performance characteristics could still change as Anthropic continues developing it, unlike Opus 5's standard mode.

Claude Opus 5's API rejects this specific combination outright: setting thinking: {"type": "disabled"} together with effort xhigh or max returns an HTTP 400 error rather than silently ignoring the disable request or falling back to some default behavior.

This is a deliberate restriction - Anthropic requires thinking to stay enabled at those two highest effort levels, so the only ways to actually disable thinking are to also lower the effort level to high or below, or to leave thinking enabled and rely on the effort setting itself to control cost instead.

For teams migrating from Opus 4.8, where no such restriction existed, this specific combination is one of the concrete, testable request patterns worth checking explicitly during migration, since it's a clear pass/fail case rather than a subtler behavioral shift.

Claude Opus 4.8 supported Priority Tier, a paid enterprise feature some teams relied on for their traffic, while Claude Opus 5 does not support Priority Tier at all.

This is a concrete capability regression to account for during migration - enterprise teams paying for Priority Tier on Opus 4.8 traffic need to make an explicit decision about how to handle that traffic before or during a move to Opus 5, rather than assuming the feature carries over automatically.

Unlike most other migration considerations, which are behavioral adjustments that can be worked around with prompting or configuration changes, this one is a hard capability gap with no equivalent replacement currently available on Opus 5.

Claude Opus 5 already performs self-verification on its own by default, more consistently than Opus 4.8 did, so an explicit "double-check your work" instruction that was useful on the older model now layers on top of behavior the model is already doing unprompted.

The practical result is over-verification: the model spends additional tokens re-checking work it had already effectively verified, without a corresponding improvement in output quality to justify that extra cost.

Anthropic's own prompting guidance frames the fix simply - delete the blanket global instruction, and only add a scoped, specific verification instruction back in if a particular step genuinely still needs one, rather than applying a general rule that made more sense on a model without this built-in behavior.

Because the minimum cacheable prompt length dropped from 1,024 tokens on Opus 4.8 to 512 tokens on Opus 5, system prompts that fell in that 512-1,024 token range - too short to cache before - now qualify for caching automatically.

This applies with no code changes required: simply moving the same request to Opus 5 makes previously-uncacheable short prompts eligible, rather than needing any explicit reconfiguration.

For high-volume agentic workloads that reuse a relatively short, consistent system prompt across many requests, this can translate into a real, if incidental, cost reduction purely as a side effect of migrating models rather than any deliberate prompt engineering.

Both models use the effort parameter, alongside query complexity, to calibrate how much thinking actually happens on a given turn, and both share the same high default effort level.

The key structural difference is that Opus 5 allows thinking to be explicitly disabled, at effort high or below, while Fable 5 and Mythos 5 always run with adaptive thinking on and have no disabled mode at any effort level.

So for Fable 5, effort and query complexity are the only levers that shape how much thinking happens, since turning thinking off entirely simply isn't an available option the way it is, with restrictions, on Opus 5.

Scope a verification instruction to a specific step, rather than removing all verification instructions outright, when that particular step carries meaningfully higher stakes or risk than the rest of the task - somewhere a mistake would be especially costly to miss even given Opus 5's generally strong built-in self-checking.

This differs from a blanket, task-wide "double-check everything" instruction, which is the pattern Anthropic specifically advises against, since it compounds with the model's default behavior across every step rather than targeting the one place extra scrutiny is actually warranted.

In practice, this means identifying the specific high-stakes step - a financial calculation, a destructive action, a security-relevant check - and attaching an explicit verification instruction only there, while trusting Opus 5's default self-verification for the rest of the task.

Anthropic's Opus releases accelerated notably over this period: Opus 4.1 shipped around August 2025, Opus 4.5 followed roughly three months later in November 2025, and subsequent versions, 4.6, 4.7, 4.8, arrived in progressively tighter succession, with Opus 4.8 landing May 28, 2026, just 41 days after Opus 4.7.

Opus 5 continued that tightened cadence, arriving roughly eight weeks after Opus 4.8 on July 24, 2026, and was itself described as Anthropic's fourth Claude 5-generation model release in under two months, following Sonnet 5 and the Mythos-tier Fable 5 and Mythos 5 models.

This reflects a broader shift Anthropic has been making from infrequent, large launches toward more frequent, incremental improvements to capability, cost, and speed delivered in quicker succession.

Effort doesn't only scale how much internal reasoning Claude Opus 5 performs before answering - it governs essentially every token the model produces during a turn, including the depth of tool calls it makes and how much elaboration goes into constructing tool arguments, not just the thinking block itself.

This means two requests with identical prompts but different effort levels can differ meaningfully not just in how thorough the final answer feels, but in how many tool calls get made, how carefully each call is reasoned through before being issued, and how much total token volume the whole turn consumes end to end.

Practically, this is why Anthropic recommends running an effort sweep against your own evaluation suite when tuning Opus 5, rather than assuming a single effort level that worked well on a narrower notion of reasoning depth will also be the right choice once its effect on tool-calling behavior and total token cost is accounted for.

It also explains why effort and response verbosity are separate levers that don't move together: since effort's influence is concentrated in thinking, tool calls, and argument construction rather than the final visible reply's wording, controlling the two independently - effort via the parameter, length via explicit prompt instructions - is necessary to get both dialed in correctly.

Start by testing existing Opus 4.8 prompts as-is, since Anthropic reports Opus 5 performs well out of the box on them - this isn't a full prompt rewrite, but a targeted tuning pass on a specific, known set of behaviors.

Add an explicit conciseness instruction if your product depends on a particular response length or verbosity, since Opus 5's default responses run longer and effort doesn't reliably control that - something like asking for focused, concise answers with minimal caveats, spending most of the response on the main point.

Remove blanket "double-check your work" verification instructions to avoid over-verification, and instead scope any remaining verification instruction to the specific steps that genuinely warrant it.

For agentic workflows, add explicit guidance on narration cadence - one sentence before the first tool call, brief updates only on meaningful changes, and a final message leading with the outcome - and review subagent delegation logic against Opus 5's more willing delegation behavior, since orchestration tuned for Opus 4.8's patterns may behave differently now.

First confirm whether thinking is running on the affected request - since Opus 5 runs thinking by default, and thinking tokens share the same max_tokens budget as the visible reply, this is the most likely cause if the same request and max_tokens value worked fine on Opus 4.8 without truncation.

Check the actual token breakdown in the response, thinking tokens versus reply tokens, to confirm how much of the budget thinking is actually consuming for this specific type of request, rather than assuming a fixed proportion, since thinking depth scales with the configured effort level and the task's complexity.

If confirmed, the fix is to raise max_tokens to account for the added thinking budget, lower the effort level so thinking consumes less of the shared budget, or, if thinking genuinely isn't needed for this request type and effort is high or below, explicitly disable thinking.

Going forward, it's worth explicitly testing max_tokens headroom as part of migration validation for any request pattern that was previously tuned assuming the full budget went to the visible reply, rather than discovering the issue only when it surfaces in production as unexpectedly truncated answers.

In a standard agentic session prior to this feature, the tool list provided in a request needed to stay fixed for the life of the conversation - changing which tools were available on a later turn would invalidate the prompt cache built up from earlier turns, forcing an expensive cache miss.

With mid-conversation tool changes enabled, a developer can submit a later turn in the same conversation with a modified tool list, adding a newly relevant tool or removing one that's no longer needed for the current stage of the task, and the underlying prompt cache from earlier turns remains valid rather than being invalidated by that change.

This makes it practical to structure an agent's capabilities progressively - for instance, starting a task with only read-only, investigative tools available, then adding a deployment or write-access tool only once the task reaches a stage where that's appropriate - without paying a caching penalty each time the tool set changes.

This is currently a beta capability accessed through a specific mechanism rather than a universally-on default, so it needs to be explicitly enabled to take effect.

flowchart LR
  A[Turn 1: initial tool list] --> B[Prompt cache built from turn 1]
  B --> C[Turn 2: modified tool list submitted]
  C --> D{Mid-conversation tool changes enabled?}
  D -- Yes --> E[Cache from turn 1 remains valid]
  D -- No --> F[Cache invalidated, forcing a cache miss]

Rather than issuing a blanket "don't narrate" instruction, which Anthropic's guidance suggests is less effective, describe the specific cadence you actually want: for example, one sentence before the first tool call to set expectations, brief updates only when something important surfaces or the plan changes direction, and a final message that leads with the outcome rather than a recap of the process.

Positive, concrete descriptions of desired communication style tend to outperform purely negative instructions listing what not to do, a pattern that applies to narration specifically but also reflects a more general prompting principle for Opus 5.

For user-facing agentic products specifically, it's worth testing this narration guidance against real user sessions rather than just isolated test prompts, since the right cadence often depends on how much visibility into the process your specific users actually want during a multi-step task.

Combine narration-cadence guidance with the separate conciseness instructions needed to control overall response length, since the two address related but distinct behavioral shifts, more frequent narration and longer default responses, that don't automatically get fixed by tuning just one of them.

For a task genuinely at this scale and duration, Anthropic's own guidance points toward Fable 5 as the recommended choice, since it explicitly calls out Fable 5 as better suited to more advanced projects, including ones a model may handle autonomously over multiple days, while positioning Opus 5 as the everyday, cost-efficient option for the bulk of typical work.

The practical reasoning is that a long-horizon autonomous task compounds small capability gaps over many sequential decisions - a model that's close to frontier-level on a single-turn benchmark can diverge more meaningfully from the frontier model's reliability once errors have many opportunities to accumulate across days of unsupervised execution.

That said, the right choice isn't purely about model capability in isolation - it's worth weighing Opus 5's roughly half price against Fable 5's stronger track record on this specific use case, and considering whether the task can be broken into shorter, more supervised checkpoints where Opus 5's cost efficiency becomes the more decisive factor instead.

In short: default to Fable 5 specifically for genuinely long, unsupervised, high-stakes autonomous runs where reliability compounding matters most, and default to Opus 5 for everything else, including most agentic coding work that fits within a normal working session.

Check the effort level configured on the failing request first - this specific error occurs when thinking: {"type": "disabled"} is combined with effort xhigh or max, since Claude Opus 5 requires thinking to remain enabled at those two highest effort levels.

If the workload genuinely needs xhigh or max effort for its capability requirements, the fix is to remove the disabled-thinking configuration and let thinking run, using the effort level itself, rather than disabling thinking outright, to manage cost and depth.

If disabling thinking is the actual priority for this specific request, whether for latency, cost, or output-format requirements, the fix is instead to lower the effort level to high or below, which is the range where explicitly disabling thinking remains a supported configuration.

Because this is a hard, deterministic restriction rather than a probabilistic behavior, it's straightforward to catch in automated testing by explicitly exercising this exact combination as part of migration or regression tests, rather than only discovering it when it surfaces unexpectedly in production traffic.

Fast mode trades roughly double the per-token price, $10 input/$50 output versus $5/$25 on standard, for approximately 2.5 times faster output, a meaningfully different cost-speed tradeoff curve than simply raising or lowering the effort parameter on standard mode.

Because it's currently a research preview available only through the Claude API, adopting it also means accepting platform lock-in for that specific workload - it isn't an option for teams routing traffic through Amazon Bedrock, Google Cloud, or Microsoft Foundry, so a multi-platform deployment can't apply it uniformly across all its Opus 5 traffic.

As a research preview rather than a generally-available feature, its behavior, pricing, and availability carry more risk of changing over time than standard mode's, worth weighing for any production workload planning to depend on it long-term rather than for exploratory or lower-stakes use.

The practical decision point is usually workload-specific: latency-sensitive interactive loops, like live drafting or exploration, benefit most from the speed gain, while high-volume, cost-sensitive batch-style workloads are generally better served by standard mode, possibly combined with a lower effort level instead of a faster but pricier mode.

Take advantage of the lower prompt-caching minimum, 512 tokens versus 1,024 on Opus 4.8, by checking whether previously-uncacheable short system prompts now qualify automatically, since this can reduce cost with literally no code changes for workloads that reuse a consistent, short system prompt at high volume.

Run an effort sweep against your own evaluation suite rather than carrying over an Opus 4.8 effort setting unchanged, since effort levels were recalibrated between the two models - a setting that was cost-optimal on Opus 4.8 may not be the cost-optimal choice on Opus 5 even if you keep the same label.

Remove or scope down blanket verification instructions that would otherwise compound with Opus 5's stronger built-in self-verification, since over-verification directly translates to wasted tokens at scale, an effect that's proportionally larger the higher the request volume.

Reserve Fast mode specifically for the subset of requests where its speed genuinely matters, rather than applying it broadly, since its roughly double per-token price makes it a poor default for cost-sensitive, high-volume traffic compared to tuning effort on standard mode instead.

With thinking disabled, at effort high or below, Claude Opus 5 can occasionally write what should have been a structured tool call directly into its user-facing text output instead of emitting a proper tool_use block - the turn completes normally, but the intended tool call never actually runs.

This is particularly disruptive in agentic loops, since the leaked, non-functional text describing the intended tool call stays in the conversation history going forward, meaning later turns are also affected by that malformed artifact rather than the issue being contained to a single turn.

This behavior shows up most commonly on tool-heavy workloads, such as search-driven tasks, where the model is frequently deciding whether and how to invoke a tool as part of its normal flow.

Anthropic's primary recommended mitigation is to avoid disabling thinking in the first place and instead manage token cost through a lower effort level, since for most tasks, thinking enabled at low effort tends to outperform thinking disabled at a similar overall cost - addressing the root cause rather than working around the artifact after the fact.

Start by auditing the prompt for verification instructions carried over from an Opus 4.8 setup - a blanket "double-check your work" or similar global instruction is the most common source, since it compounds with Opus 5's stronger built-in self-verification rather than adding proportional value.

Check whether the extra verification is showing up as genuinely repeated re-checking of the same work, versus the model simply producing a somewhat longer response by default, since these are related but distinguishable symptoms - Opus 5's baseline verbosity is also higher than Opus 4.8's, and it's worth isolating which effect is actually driving the token cost you're trying to reduce.

If a global verification instruction is confirmed as the cause, remove it, and instead scope any remaining explicit verification instruction narrowly to the specific step or steps where a mistake would be especially costly, rather than applying it across the whole task.

Validate the fix by comparing token usage and output quality before and after the change on your own evaluation set, since the goal is confirming that removing the blanket instruction reduces cost without meaningfully regressing correctness on the steps that matter, not just assuming the fix worked based on the prompting guidance alone.

Opus 5's design, approaching Fable 5's intelligence on many tasks at roughly half the price, alongside a tunable effort parameter that lets teams trade capability for cost on a sliding scale, reflects a strategic shift Anthropic has been signaling: that competition in AI is moving from purely chasing raw peak capability toward optimizing the economics of everyday, sustained use.

Positioning Opus 5 as the default for Claude Max and the strongest option on Claude Pro, while still reserving Fable 5 for the most demanding and longest-running work, creates a deliberate two-tier structure: a frontier model for pushing capability boundaries, and a close-but-cheaper everyday model for the volume of typical usage that doesn't need that ceiling.

This mirrors the pattern seen in Opus's own recent release cadence - increasingly frequent, incremental releases (4.5, 4.6, 4.7, 4.8, then 5) rather than infrequent large leaps - suggesting Anthropic is optimizing for continuous refinement of the cost-capability curve across its whole lineup, not just periodic jumps in what's possible at the very top end.

The self-described emphasis on Opus 5 being among the most aligned Opus models and less susceptible to misuse also signals that this strategy isn't purely about cost and speed - safety and alignment refinement are being positioned as part of the same everyday-model value proposition, not treated as a separate concern reserved only for the frontier-tier Fable 5 model.