Claude Opus 5: A Technical Guide for Teams Evaluating It in Production

Claude Opus 5 is Anthropic's Opus-tier model for complex agentic coding and enterprise work, released on July 24, 2026. It runs on a 1M token context window, supports a five-level effort ladder from low to max, and is priced at $5 per million input tokens and $25 per million output tokens, the same rate as its predecessor Claude Opus 4.8. This guide covers what changed, how it compares to GPT, Gemini, and Kimi K3, what it costs to run in production, and where it is the wrong tool.

The short version for engineering teams: the model did not get more expensive, it got more capable per dollar, and the API contract changed in two ways that will break some existing integrations. Both are documented below.

What is Claude Opus 5?

Claude Opus 5 is the fourth model in Anthropic's Claude 5 generation, following Mythos 5, Fable 5, and Sonnet 5. Anthropic positions it as the model most teams should reach for by default, describing it as coming close to the frontier intelligence of Claude Fable 5 at half the price.

The Claude lineup is now split across five tiers, and understanding the split matters more than it did in previous generations, because the tiers are no longer a simple ladder of size and price.

How Opus 5 differs from Sonnet and Haiku

The three tiers are not the same model at different sizes. They are different points on a capability, latency, and cost curve.

• Haiku 4.5 is for classification, extraction, routing, and other high-volume tasks where a wrong answer is cheap to detect and correct. It has a 200k context window, not 1M.
• Sonnet 5 is the volume workhorse: strong instruction following and writing, 1M context, and roughly 40% of Opus 5's per-token cost at standard rates. Most production traffic in a well-designed system should land here.
• Opus 5 is for the subset of work where the model has to hold a plan across many steps: multi-file refactors, root-cause debugging, long-horizon agent runs, and analysis that has to stay correct across hundreds of thousands of tokens.
• Fable 5 remains above Opus 5 in raw capability, but carries safety classifiers that can decline requests, plus 30-day data retention. Opus 5 has no data retention requirement for general access.

The practical rule: route by task difficulty, not by model prestige. Teams that put Opus 5 behind every request pay Opus prices for work Sonnet 5 completes correctly.

What is new in Claude Opus 5?

Opus 5 introduces four new platform capabilities and two behavior changes that affect existing code. Anthropic describes it as a step-change improvement over Opus 4.8 rather than an incremental one, with the largest gains in deep reasoning, agentic and long-horizon tasks, and test-time compute scaling.

The full effort ladder, including max

The effort parameter controls how many tokens the model spends on a response. Opus 5 supports the complete ladder: low, medium, high, xhigh, and max. No beta header is required. On the Claude API and Claude Code, effort defaults to high.

This is the most important cost lever available to an engineering team. Anthropic reports that low and medium effort on Opus 5 produce strong quality at a fraction of the tokens and latency of higher settings, and perform above the same settings on prior Opus models. For coding and agentic work, xhigh is the recommended starting point. If you carried effort defaults over from Opus 4.8, re-run a sweep against your own evaluation set: the difference between medium and max on a high-volume endpoint is a material line item.

client.messages.create(
    model="claude-opus-5",
    max_tokens=64000,
    output_config={"effort": "max"},
    messages=[{"role": "user", "content": "..."}],
)

Thinking is on by default

On Opus 4.8, requests ran without thinking unless you explicitly set thinking: {"type": "adaptive"}. On Opus 5, the same request runs with thinking on. The model decides when and how deeply to think on each turn, and effort is the control for depth.

Because max_tokens is a hard ceiling on total output, thinking tokens plus response text, any workload that ran without thinking on Opus 4.8 needs its max_tokens revisited.

Breaking change: disabling thinking requires effort high or below

Setting thinking: {"type": "disabled"} with effort xhigh or max returns a 400 error on Opus 5. This is enforced per request and is generally available behavior, not a beta. Two options: keep thinking disabled and drop effort to high or below, or keep the effort level and remove the thinking field.

Anthropic also documents two artifacts that appear when thinking is disabled: the model can write a tool call into its visible text instead of emitting a structured tool_use block, and it can emit internal XML tags into visible output. The documented mitigation is to keep thinking enabled and control cost through lower effort instead.

Mid-conversation tool changes

In beta, with the mid-conversation-tool-changes-2026-07-01 header, developers can add or remove tools between turns without invalidating the prompt cache. Previously, a fixed tool list had to be resent for the life of a session.

For agent architectures this removes a real constraint. A support agent can start with a read-only tool set, gain write tools after identity verification, and drop them again afterward, without paying to rewrite the cache on every transition.

Automatic fallbacks on the API

The fallbacks parameter gained a "default" mode that applies Anthropic's recommended fallback models by refusal category rather than a list you maintain. Requests flagged by safety classifiers route to another model instead of being blocked. In Claude.ai, Claude Code, and Claude Cowork, flagged requests fall back to Opus 4.8 by default.

Lower prompt cache minimum

The minimum cacheable prompt on Opus 5 is 512 tokens, down from 1,024 on Opus 4.8. Short system prompts that were previously uncacheable now create cache entries with no code changes. Cache hits cost 10% of the base input rate, so this is a free saving for products with short, stable preambles.

Fast mode

Fast mode, in research preview, runs Opus 5 at roughly 2.5 times default speed, priced at $10 per million input tokens and $50 per million output tokens. It is available on the Claude API only, not on Amazon Bedrock, Google Cloud, or Microsoft Foundry, and it cannot be combined with the Batch API.

Behavior changes you will notice without changing code

• Default responses and written deliverables run longer than on Opus 4.8.
• In agentic sessions the model narrates progress more often.
• It delegates to subagents more readily.
• It verifies its own work without being told to.

That last point has a counterintuitive consequence. Anthropic's guidance is to remove verification instructions carried over from earlier models, such as "include a final verification step" or "use a subagent to verify". On Opus 5 they compound with the model's own behavior and cause over-verification, burning tokens with no quality gain.

Architecture and how the model works

Anthropic has not published parameter counts, training compute, or architectural details for Opus 5, and no credible third-party figure exists. What is documented is the operational surface, which is what matters for system design.

Context

Opus 5 has a 1M token context window as both the default and the maximum, with no smaller variant and no long-context price premium: a 900k-token request is billed at the same per-token rate as a 9k-token request. Anthropic states that instruction following, tool calling, and reasoning stay consistent throughout the window. Everything counts toward it, including system prompt, tool definitions, history, and thinking blocks from previous turns, which the API retains by default and bills as input tokens on later turns.

Reasoning as a budget, not a mode

The mental model that works best in production is that effort buys test-time compute. Anthropic lists "test-time compute scaling" as one of the model's largest gains, meaning additional effort converts into better results up to the max level rather than plateauing early.

This gives an architecture two independent dials instead of one. Model choice sets the capability ceiling. Effort sets how much compute is spent approaching that ceiling on a given request. A well-instrumented system varies effort per route: low for classification, medium for routine edits, xhigh for a multi-file refactor.

Planning and execution

Anthropic's prompting guidance for Opus 5 states the model performs best on difficult coding work when given the complete task specification up front and then left to run, rather than being fed the task in fragments. It completes full tasks rather than leaving stubs or placeholders.

The corresponding risk is scope expansion. The same guidance notes the model can add steps that were not requested, and recommends constraining scope explicitly for narrow tasks.

Multi-agent coordination

Opus 5 coordinates teams of subagents with effective writer-verifier patterns and few cases of agents overwriting each other's work. Because it delegates more readily than prior models, Anthropic recommends capping delegation for cost-sensitive workloads, either through prompt guidance or deterministic limits in the harness.

Claude Opus 5 benchmarks

Two categories of numbers exist for Opus 5, and they should not be read the same way.

Numbers published by Anthropic

The following are from Anthropic's own launch announcement and were produced by Anthropic, not by an independent evaluator.

• Frontier-Bench v0.1: Opus 5 surpasses all other models and more than doubles Opus 4.8's performance at a lower cost per task.
• CursorBench 3.2: at max effort, within 0.5% of Fable 5's peak score at half the cost per task.
• ARC-AGI 3 (novel problem solving): roughly three times the score of the next-best model.
• Zapier AutomationBench (end-to-end business tasks): pass rate around 1.5 times the next-best model at the same cost per task. Even at its lowest effort setting, it passes more tasks than any other model.
• OSWorld 2.0 (computer use): outperforms every other model at any given cost, surpassing Fable 5's best result at just over a third of the cost.
• Life sciences: 10.2 percentage points above Opus 4.8 on inferring molecular structures from spectroscopy data, and 7.7 points higher on protein variant effect prediction, on Anthropic's internal benchmarks.
• Alignment: a score of 2.3 on Anthropic's automated behavioral audit of misaligned behavior, the lowest of its recent models.

Independent numbers

Artificial Analysis, which runs its own evaluations like-for-like across models, placed Claude Opus 5 at adaptive reasoning and max effort at the top of its Intelligence Index at launch. The Index is a weighted average across four equally weighted categories: agents, coding, general capability, and scientific reasoning.

Gemini 3.1 Pro sits outside this particular snapshot at the top of the index but remains the strongest value proposition on price among the closed models, at $2 / $12 per MTok with a 1M context window and native audio and video input. Alibaba's Qwen 3.8 Max, a 2.4-trillion-parameter model, entered public preview in July 2026 and does not yet have comparable published independent scores.

How to read these tables

Composite index scores compress dozens of tasks into one number. That is useful for a first pass and misleading for a purchasing decision. Three specific cautions:

1. Effort settings change the result. Opus 5's 61 was measured at max effort. The same model at medium effort scores lower and costs far less. A model comparison that does not state effort settings is not comparable.
2. Cost per task beats price per token. A model that is 30% cheaper per token but uses three times the tokens to finish is more expensive. Artificial Analysis reports cost per task alongside score for this reason, and Anthropic's own charts plot performance against cost rather than in isolation.
3. Benchmarks measure benchmark tasks. Frontier-Bench and CursorBench correlate with real software engineering work, but they do not contain your codebase, your test suite, or your tooling. The only number that decides a migration is the pass rate on your own evaluation set.

Claude Opus 5 vs GPT-5.6

GPT-5.6 Sol is OpenAI's flagship in the GPT-5.6 family, priced at $5 per million input tokens and $30 per million output tokens, with a 1,050,000-token context window and 128k max output. One structural detail matters for long-context work: requests above 272,000 input tokens are billed at higher rates for the entire request, roughly double on input. Opus 5 has no such tier.

When to choose Claude Opus 5

• Long-horizon agentic coding: multi-file features, large refactors, work that has to survive many tool-use loops without drifting.
• Code review and bug finding, where Anthropic reports high precision and recall per pass with few false positives, holding accuracy at lower effort settings.
• Workloads that regularly exceed 272k input tokens, where flat pricing across the full window changes the economics.
• Regulated environments that need no data retention on general access.

When to choose GPT-5.6

• Existing investment in the OpenAI Responses API, Codex, or the broader OpenAI tooling ecosystem.
• Workloads where the cheaper family members, Terra at $2.50 / $15 and Luna at $1 / $6, cover most traffic and only a fraction needs the flagship.
• Teams standardized on OpenAI-compatible interfaces across many providers, where the SDK surface is a real switching cost.

Both models are close enough on published composites that the decision should be made on your own evaluation harness, on integration cost, and on the shape of your context usage, not on a one-point difference in an index.

Claude Opus 5 vs Gemini

Where Claude Opus 5 leads

• Agentic coding and long-horizon tool use, where Anthropic's benchmark set and third-party harnesses such as Cursor and Devin both report Opus-class strength.
• Enterprise deployment surface: the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, with a documented data residency parameter.
• Granular cost control through the five-level effort ladder rather than a three-level thinking control.

Where Gemini retains advantages

• Price. Gemini 3.1 Pro at $2 / $12 per MTok is substantially cheaper per token than Opus 5, and Gemini 3.6 Flash, launched July 21, 2026, at $1.50 / $7.50 targets coding workloads at lower cost still. Note that Gemini 3.1 Pro's rate doubles above a 200k-token context, unlike Opus 5's flat pricing.
• Native multimodality. Gemini Pro models accept text, images, audio, and video. Claude models accept text and image input only. For workloads built on video or audio ingestion this is not a close comparison.
• Google Cloud integration. For organizations already standardized on Vertex AI, BigQuery, and Workspace, the integration path is shorter.

Google's next Pro-tier release, Gemini 3.5 Pro, was announced in May 2026 but has slipped, with reporting in July 2026 indicating it is months behind schedule. Teams planning around it should treat the timeline as unconfirmed.

Claude Opus 5 vs Kimi K3

Kimi K3, released by Moonshot AI on July 16, 2026, is the most credible open-weight challenger in this comparison: a mixture-of-experts model with roughly 2.8 trillion total parameters, of which 16 of 896 experts activate per token, a 1M token context window, and native visual understanding.

What the open weights actually buy

Open weights matter when you need data to stay inside your own perimeter, when you need to fine-tune, or when you want insurance against a vendor's pricing or availability decisions. K3 delivers all three in principle. In practice, a 2.8T mixture-of-experts model with a 64-accelerator floor is not something most teams will self-host. The realistic K3 deployment for the majority of buyers is a hosted API, which converts the open-weight advantage into an option that is held rather than exercised.

Two further considerations: independent testing found K3's factual accuracy improved over K2.6 while its hallucination rate also rose, which argues for verification layers on factual workloads, and "open weights" no longer implies "cheap", since $3 / $15 sits in the same band as Claude Sonnet 5's standard rate.

Where K3 is the better choice

• Front-end and visual coding work, where K3 took the top position on Arena's Frontend Code leaderboard at launch.
• Organizations with a hard requirement that weights run on infrastructure they control.
• Cost-sensitive agentic workloads where measured cost per task, not list price, favors K3 on your evaluation set.

Claude Opus 5 in Claude Code

Claude Code is Anthropic's agentic coding tool, available in the terminal, in IDEs, and in the desktop and mobile apps. Opus 5 is available there at launch, with effort defaulting to high. It also shipped same-day into third-party surfaces including GitHub Copilot, where it is available on Copilot Pro+, Max, Business, and Enterprise tiers.

Where it earns its cost

• Multi-file feature work. The model completes tasks end to end rather than leaving stubs, which removes the review cycle spent hunting for placeholders.
• Refactoring. Large mechanical changes across a repository, where the 1M context window means the model can hold the call graph rather than guess at it.
• Debugging and root-cause analysis. Cognition reported particular strength on difficult debugging inside Devin. Anthropic describes the model finding a root cause in an open-source package manager and fixing an edge case the community patch had missed.
• Code review. High findings-per-pass with few false positives, and accuracy that holds at lower effort, which supports a cheap pass at commit time and a thorough pass before merge. Caveat from Anthropic's prompting guide: instructions like "only report high-severity issues" are followed literally and suppress findings. Ask for everything and filter in a second pass.
• Test generation. The model builds its own verification scaffolding when none exists, including test harnesses for systems with no live data source to validate against.
• Documentation. Longer default deliverables help here, though length calibration should be prompted explicitly.

Why Claude remains a default for software engineering

The reason is not any single benchmark. It is that the model behaves predictably inside an agent loop: it verifies its own work, it recovers from errors without intervention, it stays on task through extended tool-use sequences, and its variance run to run is low enough that a team can build process around it. Lovable reported a 22% improvement on its hardest agentic coding tasks over Opus 4.7, and specifically called out reduced run-to-run variance as the property that mattered for its users.

Claude Opus 5 API: pricing, integration, and best practices

Pricing

The full 1M context window is billed at standard rates. Caching and batch discounts apply across the whole window, and they stack: batch plus cache hits is the cheapest way to run large repeated-context jobs. US-only inference through the inference_geo parameter applies a 1.1x multiplier across all token categories.

One accounting note that catches teams migrating from older models: Claude 4.7 and later use a newer tokenizer that produces roughly 30% more tokens for the same text. Do not compare token counts measured against Sonnet 4.6 or earlier against Opus 5 invoices.

Availability and SDKs

Opus 5 is available as claude-opus-5 on the Claude API and Google Cloud, as anthropic.claude-opus-5 on Amazon Bedrock, and through Claude in Microsoft Foundry. Official SDKs cover Python, TypeScript, C#, Go, Java, PHP, and Ruby. Model IDs in this generation use a dateless format that is still a pinned snapshot, not a moving pointer.

Migration from Opus 4.8

1. Change the model ID from claude-opus-4-8 to claude-opus-5.
2. Revisit max_tokens, since thinking tokens now count against it by default.
3. Audit any request that sets thinking: {"type": "disabled"} together with xhigh or max effort. These now return a 400 error.
4. Remove verification and re-check instructions from prompts, which cause over-verification.
5. Re-run an effort sweep on your evaluation set instead of inheriting Opus 4.8 defaults.
6. Add explicit length and scope constraints if your product surfaces model output directly to users.

Production best practices

• Cache aggressively. A cache hit costs 10% of base input. With the 512-token minimum, most system prompts now qualify.
• Route by difficulty. Send classification and extraction to Haiku 4.5, routine generation to Sonnet 5, and reserve Opus 5 for work that genuinely needs it.
• Use the Batch API for anything not latency-bound. Half price, and it stacks with caching.
• Cap subagent spawning either in the prompt or deterministically in the harness. Delegation pays on large independent tracks of work and multiplies cost on small ones.
• Prompt for conciseness explicitly. Effort controls thinking volume, not visible response length. If you need short answers, ask for them.
• Instrument cost per successful task, not cost per call. That is the number that tells you whether an effort change or a model change was an improvement.

Enterprise use cases

• AI agents. Long-running agents that hold a goal across many tool calls. Mid-conversation tool changes make progressive permission models practical, and automatic fallbacks reduce the blast radius of a classifier trigger in a production path.
• Customer support. Opus 5 is the wrong model for the bulk of support volume, which Haiku 4.5 or Sonnet 5 handles far cheaper. It fits the escalation tier: policy reasoning, multi-system lookups, or responses with legal or financial consequence.
• Document processing. The 1M window with flat pricing suits contract sets, filings, and technical documentation. Box reported an 8% improvement over Opus 4.8 on specialized enterprise content, with larger gains in data analysis and due diligence.
• Internal knowledge and enterprise search. Consistent behavior across the full window reduces the need for aggressive chunking. Cache the corpus preamble and pay hit rates on repeat queries.
• Workflow automation. Zapier reported that Opus 5 topped its AutomationBench leaderboard without spending more tokens than prior Claude models, completing an end-to-end churn-prevention sequence that previous models did not pass.
• Product development. Discovery synthesis, specification drafting, prototype generation, and design-to-code work, where Anthropic reports improved visual understanding and frontend replication when the model can inspect and verify its own output.
• Software engineering. Code review, refactoring, migrations, test generation, and incident debugging, covered in the Claude Code section above.

Limitations and when not to use Claude Opus 5

An honest evaluation has to include the cases where this model is the wrong answer.

• Cost at volume. At $5 / $25 per MTok, high-frequency workloads that Sonnet 5 or Haiku 4.5 handle correctly should not be on Opus 5. The tokenizer change since Claude 4.7 also means more billed tokens for the same text than older-generation comparisons suggest.
• Verbosity. Default responses and written deliverables run longer than on Opus 4.8, and effort does not reliably shorten visible output. Products with tight UI constraints need explicit length instructions.
• Scope expansion. The model can widen a task beyond what was asked. For narrowly defined jobs this needs to be constrained in the prompt.
• Cybersecurity restrictions. Opus 5's classifiers allow finding vulnerabilities in source code but block binary-based vulnerability scanning, penetration testing, and exploit generation. Security teams with legitimate offensive needs must go through Anthropic's Cyber Verification Program.
• Not the ceiling. Fable 5 remains more capable, and Mythos 5 is ahead on both biology research and offensive cybersecurity. Opus 5 also remains limited on long-running autonomous research tasks in biology.
• API dependency. There are no open weights, no self-hosting, and no offline deployment. Availability, pricing, and model lifecycle are Anthropic's decisions. Organizations with air-gapped requirements need an open-weight path such as DeepSeek V4 Pro under MIT, or Kimi K3 once its weights land.
• Thinking-disabled artifacts. Integrations that must run with thinking off should expect occasional leaked tool calls and internal tags, and plan mitigations.

Is Claude Opus 5 worth using?

• Startups: usually yes, but not for everything. Opus 5 for the hard 10%, Sonnet 5 for the rest, effort tuned per route. Small teams gain more from unattended completion of large tasks, because engineering time is the scarcer input.
• Software houses and consultancies: yes. Delivery work is dominated by unfamiliar codebases, migrations, and refactors, which is the profile Opus 5 targets. Low run-to-run variance matters more than a marginal benchmark edge, because estimates depend on predictability.
• Large enterprises: yes for the reasoning tier, with governance. Availability across AWS, Google Cloud, and Microsoft Foundry, no data retention on general access, and a data residency parameter address most procurement objections. Set budget guardrails at the routing layer, not per team.
• Engineering teams: yes, provided you build an evaluation set before migrating. Every claim in this article is a starting hypothesis; your own pass rates and cost-per-task numbers decide.
• SaaS products: conditionally. Opus rates only work behind a feature priced to match, or on a low-frequency, high-value path. Model unit economics per feature, not per model.

Frequently asked questions

What is Claude Opus 5?

Claude Opus 5 is Anthropic's Opus-tier large language model, released July 24, 2026, built for complex agentic coding and enterprise work. It has a 1M token context window, 128k max output, thinking on by default, and a five-level effort ladder from low to max.

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Batch processing halves both rates. Cache hits cost $0.50 per million tokens. Fast mode is $10 / $50.

Does Claude Opus 5 have an API?

Yes. The model ID is claude-opus-5 on the Claude API and Google Cloud, and anthropic.claude-opus-5 on Amazon Bedrock. It is also available through Claude in Microsoft Foundry.

What is the Claude Opus 5 context window?

1 million tokens, which is both the default and the maximum. There is no smaller variant and no long-context price premium.

Is Claude Opus 5 better than GPT-5.6?

On the Artificial Analysis Intelligence Index at launch, Opus 5 at max effort scored 61 against GPT-5.6 Sol's 59, and Opus 5's output tokens are cheaper at $25 against $30 per million. GPT-5.6 also applies higher rates to requests above 272k input tokens. The gap is small enough that the decision should rest on your own evaluations and integration costs.

Is Claude Opus 5 better for coding?

Anthropic reports state-of-the-art results on Frontier-Bench v0.1 and near-parity with Fable 5 on CursorBench 3.2 at half the cost per task. Independent harness vendors including Cursor, Cognition, and Lovable reported gains on their internal coding evaluations. For front-end visual work specifically, Kimi K3 led Arena's Frontend Code leaderboard at its launch.

Claude Opus 5 vs Gemini: which should I use?

Opus 5 for agentic coding, long-horizon tool use, and finer cost control through the effort ladder. Gemini 3.1 Pro for lower per-token cost at $2 / $12, native audio and video input, and deep Google Cloud integration. Note that Gemini 3.1 Pro's rate doubles above a 200k context, while Opus 5 pricing is flat across 1M.

Claude Opus 5 vs Kimi K3: which should I use?

Opus 5 leads on independent composite intelligence scores and on enterprise deployment surface. K3 is cheaper per token, ships open weights, and led the Frontend Code Arena at launch. Self-hosting K3 requires substantial infrastructure, so for most teams the open-weight advantage is optionality rather than an immediate deployment path.

Can Claude Opus 5 build AI agents?

Yes. It is explicitly positioned for long-horizon agentic tasks, supports mid-conversation tool changes without invalidating the prompt cache, coordinates subagents with writer-verifier patterns, and offers automatic model fallbacks when safety classifiers trigger.

Is Claude Opus 5 good for enterprise software?

It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, has no data retention requirement for general access, and supports US-only inference through the inference_geo parameter at a 1.1x multiplier. Anthropic also reports it as its most aligned Opus model on its internal behavioral audit.

What is the effort parameter in Claude Opus 5?

Effort controls how many tokens the model spends on a response, across five levels: low, medium, high, xhigh, and max. It defaults to high on the Claude API and Claude Code. Anthropic recommends xhigh as the starting point for coding and agentic work, and notes that low and medium produce strong quality at a fraction of the tokens and latency.

Is Claude Opus 5 open source?

No. It is a proprietary model available only through Anthropic's API and partner cloud platforms. Teams that require open weights should evaluate DeepSeek V4 Pro, released under MIT, or Kimi K3.

What are Claude Opus 5's cybersecurity restrictions?

Its classifiers permit finding vulnerabilities in source code but block binary-based vulnerability scanning, penetration testing, and exploit generation. Anthropic expects these classifiers to intervene roughly 85% less often than Fable 5's. Enterprises and researchers with legitimate needs can access a less restricted version through Anthropic's Cyber Verification Program.

Which Claude model should I use by default?

Anthropic's own guidance is to start with Opus 5 for complex agentic coding and enterprise work, and to use Fable 5 where the highest available capability is required. In production systems, most traffic belongs on Sonnet 5 or Haiku 4.5, with Opus 5 reserved for tasks whose difficulty justifies the rate.

Where this leaves engineering teams

The pattern in this release generation is that capability is no longer the scarce resource. Four models now sit within a few points of each other on independent composites, and the differences that decide a production stack are elsewhere: cost per successful task, behavior inside an agent loop, deployment surface, data handling, and how much prompt and harness work is needed to get consistent output. That shifts the problem from model selection to systems engineering. The teams getting the most out of Opus 5 have a routing layer that sends each request to the cheapest model that passes, an evaluation set that tells them when that changes, and instrumentation reporting cost per successful outcome rather than tokens consumed.

At Amplifi Labs we build that layer: AI engineering and integration, agent architectures, and the product engineering around them, from discovery through to production. If you are evaluating Claude Opus 5 against your own workload and want the evaluation harness and routing design in place before you migrate, that is the work we do.