Claude Opus 5 Ranks #1. Why Anthropic Still Gates Mythos.

Claude Opus 5 launched on July 24, 2026, and its max-effort configuration ranked #1 on the Artificial Analysis Intelligence Index when we checked. Yet Anthropic still calls Fable 5 its highest-capability generally available model, while Mythos 5—the same underlying model with certain cyber and biology classifiers removed—remains gated behind Project Glasswing.

That apparent contradiction is the real launch story. Anthropic no longer sells one universal “best Claude.” It sells a public leaderboard winner, a maximum-capability generally available tier, and a gated configuration with select classifiers removed to three different buyers. Opus 5 is the $5-input, $25-output, one-million-context option built to win the largest market: everyday serious work.

This guide covers the release, verified API pricing, specifications, access, benchmark context, and the Opus-versus-Fable decision. Last source check: July 25, 2026. Facts were checked against Anthropic's launch announcement, the Claude model overview, Anthropic's Fable and Mythos access notes, current Claude API pricing, and Artificial Analysis' live Opus 5 profile.

The real Opus 5 story: Anthropic split “best” into three products

Most Opus 5 launch posts will reproduce the same specification sheet. The strategic change is that Anthropic has separated benchmark leadership, maximum generally available capability, and classifier-gated access into different products.

There is a quieter enterprise difference most launch coverage misses. Anthropic's migration guide says Fable 5 and Mythos 5 require 30-day data retention and are not available under zero-data-retention arrangements. Opus 5 does not carry that model-specific requirement. For a regulated buyer, that can rule out the higher tier before benchmarks or token price even enter the conversation.

This is why “Anthropic's third-best model” is a strong hook but a weak fact. Fable and Mythos are not two separate intelligence ranks; they are two access and safeguard configurations of the same underlying model. Opus is a different, cheaper product that can still win a broad public index. The more accurate—and more interesting—conclusion is that Anthropic has decoupled leaderboard rank from its premium capability tier.

Opus 5 may be Anthropic's most commercially important model even if it is not its most capable one. Buyers do not purchase abstract intelligence. They purchase a model configuration, latency profile, safeguard policy, and bill. Opus 5 reaches close to Fable performance on several valuable tasks without Fable's $10-input and $50-output rates.

What is Claude Opus 5?

Claude Opus 5 is Anthropic's premium daily-driver model for complex agentic coding, long-running workflows, document creation, and professional knowledge work. It sits between Sonnet 5's speed-and-value position and Fable 5's highest-capability position. The weights and parameter count are proprietary and have not been disclosed.

What's new in Claude Opus 5 vs Opus 4.8?

Opus 5's biggest changes are thinking by default, a stronger effort-to-cost curve, smaller cacheable prompts, and better support for long-running tool workflows. The current Opus 5 developer guide documents the following:

Opus 5 is the successor most teams should evaluate, but Opus 4.8 has not disappeared. Anthropic still lists claude-opus-4-8 as active, and it remains part of the safety fallback path. Anthropic says Opus 5 will not retire before July 24, 2027.

Claude Opus 5 pricing and access

Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens at standard speed. Prompt-cache reads are 90 percent cheaper than standard input, Batch is half price, and Fast mode doubles the base rates.

There is no long-context surcharge. A 900,000-token request uses the same per-token rate as a short request. US-only inference on Anthropic-operated platforms carries a 1.1× multiplier. Fast mode is a first-party Claude API research preview; it is not available on partner cloud platforms or with the Batch API.

One cost caveat is easy to miss. Anthropic says Claude 4.7 and later models use a newer tokenizer that produces roughly 30 percent more tokens for the same text than the earlier tokenizer, depending on the workload. Opus 4.8 already uses that tokenizer, so the Opus 5 list price is directly comparable with 4.8. Teams migrating from Opus 4.6 or older should compare the invoice for a representative task, not only the price-per-million row.

Where can you access Claude Opus 5?

Opus 5 is available on Claude Pro, Max, Team, and Enterprise plans, through Claude Code and the Claude API, and across Anthropic's supported cloud platforms. It is the default model on Claude Max and the strongest model included on Claude Pro. Developers can use it through:

• Claude API: model ID claude-opus-5.
• Amazon Bedrock: anthropic.claude-opus-5 through the Messages-API Bedrock endpoint.
• Google Cloud: claude-opus-5.
• Claude Platform on AWS and Microsoft Foundry: listed as supported launch platforms.

The dateless model ID is still a pinned snapshot, not an evergreen alias. In other words, claude-opus-5 will not silently become a future Opus release. That is good for reproducibility, but it means teams still need a model-upgrade process.

How to use Claude Opus 5 in the API

The current Anthropic Python SDK accepts effort through output_config. This example starts at medium effort to control cost; omit the field to use the high-effort default.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5",
    max_tokens=4096,
    output_config={"effort": "medium"},
    messages=[{
        "role": "user",
        "content": "Review this API design for failure modes.",
    }],
)

for block in response.content:
    if block.type == "text":
        print(block.text)

Keep the API key on the server. For coding and agentic work, Anthropic recommends starting at xhigh; for most other intelligence-sensitive work, start at high. Use low or medium when your evaluations show quality holds. Reserve max for tasks that justify unconstrained token spending, and give xhigh or max enough output room—Anthropic suggests starting around 64K max_tokens.

There is a caching wrinkle: changing effort between requests changes the rendered prompt and can invalidate the cached prefix. Choose effort per workload, but keep it stable inside a long cached conversation.

Claude Opus 5 benchmarks: where it wins and what is missing

Anthropic reports state-of-the-art results on several coding, knowledge-work, and computer-use evaluations; Artificial Analysis ranked the max-effort configuration first on its composite index when checked. That is a strong launch, but not evidence that Opus 5 wins every workload.

Two details keep the headline honest. First, the Artificial Analysis result is the max-effort configuration, not the API's default high setting. Its evaluation generated 100 million output tokens and cost $3,835.51, according to the live profile. Second, Anthropic's Frontier-Bench footnote says Opus 4.8 served as the fallback when safety classifiers refused tasks for Opus 5 and Fable 5. Those rows measure deployed configurations—model, effort, harness, and fallback—not isolated weights.

That is why cost per completed task is the more useful buying signal. A model that spends twice the tokens to gain one benchmark point may be worse for production, while a lower effort setting that holds quality on your own evaluations may be the real winner. For cross-vendor context, see our current guides to GPT-5.6 Sol, Terra, and Luna and Kimi K3.

Claude Opus 5 vs Fable 5 vs Sonnet 5

Choose Opus 5 when you need serious coding and agentic capability but cannot justify Fable 5's twice-as-high token rates. Choose Fable only when the work is long, difficult, and valuable enough that your own evaluations show a measurable advantage. Choose Sonnet for high-volume or latency-sensitive work.

The strongest marketing claim Anthropic can make is not "Opus beats Fable." It is "Opus makes Fable unnecessary for more work." If Opus 5 at high or xhigh passes your acceptance criteria at half the base token rate, paying for Fable is hard to defend. If a long autonomous workflow fails on Opus and passes on Fable, the premium is easy to defend. Run that decision on representative tasks, not a global leaderboard.

The Opus 4.8 safety fallback, explained

A narrow set of cybersecurity requests flagged by Opus 5's safeguards can be routed to Opus 4.8 rather than refused outright. Anthropic enables that behavior by default in Claude.ai, Claude Code, and Claude Cowork. API customers can opt into server-side fallbacks.

That is usually a better product experience than a dead-end refusal, but it matters for audits and evaluations: two requests sent to "Opus 5" can be completed by different models. Record the returned model and fallback metadata when reproducibility or compliance matters.

Our practical recommendation for builders

1. Start at high, not max. High is the API default and the sensible baseline for difficult work. Use xhigh for serious coding and agentic tasks; move to max only when your evaluation shows a real gain.
2. Run an effort sweep. Test low, medium, high, xhigh, and max on the same representative tasks. Measure pass rate, wall-clock time, input and output tokens, tool calls, and retries.
3. Optimize completed-task cost. Token price without token volume, retries, and human correction time is not a cost model.
4. Cache stable context. The $0.50 cache-read rate can materially change long-agent economics. Keep effort fixed inside cached conversations.
5. Use Batch and Fast for opposite jobs. Batch is for asynchronous volume at half price. Fast mode is for interactive work where latency costs more than tokens.

For teams building one product rather than a model-routing layer, the durable play is to hide this complexity behind task-level evaluations and routing. That is also how Casagbic approaches app building: the user describes the outcome, while the platform handles which model and effort profile should do the work.

Frequently asked questions

Sources and methodology

We checked release, pricing, specifications, access, effort guidance, safeguards, and benchmark claims against the sources below on July 25, 2026. Vendor evaluations are labeled as self-reported; live third-party rankings are dated because they can change.

• Anthropic: Introducing Claude Opus 5 — release date, positioning, launch evaluations, safeguards, access, and Fast mode.
• Claude Platform: Models overview — model IDs, modalities, context, output, knowledge cutoff, cloud availability, and Claude-family comparison.
• Claude Platform: What's new in Claude Opus 5 — thinking behavior, cache changes, fallbacks, tool changes, and Fast mode constraints.
• Claude Platform: Migration guide — API behavior and the 30-day retention requirement that applies to Fable 5 and Mythos 5 but not Opus 5.
• Claude Platform: Effort — the five supported levels, defaults, API example, and model-specific guidance.
• Claude Platform: Pricing — base, cache, Batch, Fast, long-context, data-residency, and tokenizer details.
• Anthropic: Claude Fable 5 and Claude Mythos 5 — the shared underlying model, safeguard differences, Project Glasswing access, and Anthropic's capability-tier positioning.
• Claude Platform: Fable 5 and Mythos 5 — capability relationship, pricing, safeguards, and limited availability.
• Artificial Analysis: Claude Opus 5 profile — independent live index score, configuration, token use, and evaluation cost.

The field will move again. The durable decision is to evaluate a model on your own tasks, hold it to an acceptance test, and optimize for the total cost of work that actually passes. Want to skip the model routing and get an app built? Try it on Casagbic →