I Traced 4 Claude Opus 5 Signals. The Release Date Still Isn't Real Yet.

What Anthropic actually lists today

The official Claude model overview currently lists four main public tiers:

There is no Opus 5 row. There is no claude-opus-5 ID. There is no Opus 5 system card in Anthropic's system-card index. I keep repeating that because an API model name is one of the easiest rumors to fake. A string found in a client bundle can be a placeholder. A gateway catalog can use its own alias. A screenshot can be edited. I don't consider a model real for developers until the first-party catalog or API exposes it.

The cadence signal is real, but weaker than it looks

Anthropic has shipped Opus updates quickly:

If I mechanically apply the observed 42-73 day range to May 28, I get July 9 through August 9. That arithmetic is valid. The forecast is fragile. Three intervals are a tiny sample. More importantly, Anthropic changed the lineup on June 30 by launching Sonnet 5 and restoring Fable 5. A company doesn't have to keep shipping one product family on schedule while it is still explaining two adjacent tiers.

My actual read is:

I won't turn those labels into fake percentages. There isn't enough evidence to say "64% chance by August 9" with a straight face.

The product gap is the strongest clue

Sonnet 5 is cheap enough to be the default production model. Fable 5 is powerful enough to be the premium long-horizon model. But the price doubles between current Opus and Fable. That leaves room for a model that improves on Opus 4.8 without forcing every serious agent workload onto Fable's $10/$50 rate.

Here's the cost shape for 100 million input tokens and 20 million output tokens per month:

That $1,000 monthly gap is why I think an Opus 5 tier still makes commercial sense. If Anthropic can deliver part of Fable's agent reliability at Opus pricing, it has a clean product. If it simply renames Fable and keeps $10/$50, Opus becomes much less meaningful as a separate tier.

The $5/$25 prediction is reasonable, not confirmed

Opus 4.5 cut the tier to $5 input and $25 output per million tokens. Opus 4.6, 4.7, and 4.8 kept it. That is four consecutive versions at one price. It also fits neatly between Sonnet 5's eventual $3/$15 and Fable 5's $10/$50. So yes, if I had to build a planning scenario today, I'd use $5/$25 as the base case. But I would put UNCONFIRMED beside it in capital letters.

The same applies to these likely features:

• A 1M-token context window
• Adaptive thinking
• An effort control
• Prompt caching
• Batch pricing
• A model ID shaped like claude-opus-5

All are consistent with the current Claude family. None is an Opus 5 API fact.

Three cost scenarios I'd model before launch

I don't need fake benchmarks to prepare a migration budget. I need a few price scenarios.

1. Small production agent
Monthly volume: 10M input, 2M output.

• At $5/$25: 10 x $5 + 2 x $25 = $100/month
• At $10/$50: 10 x $10 + 2 x $50 = $200/month

The premium scenario adds $1,200 a year. On one service, that's manageable. Across 50 internal agents, it's $60,000.

2. Coding platform
Monthly volume: 100M input, 20M output.

• At $5/$25: $1,000/month
• At $10/$50: $2,000/month

Annual difference: $12,000. I would require the premium model to save more than $1,000 per month in retries, engineering review, or failed tasks before moving all traffic.

3. Cached agent context
Suppose the same coding platform has 100M input, but 80M tokens are cache hits. At current Opus 4.8 rates:

• 20M fresh x $5 = $100
• 80M cached x $0.50 = $40
• 20M output x $25 = $500
• Total = $640/month

Caching saves $360 against the uncached $1,000 bill. That's a real optimization available today. Waiting for an imaginary benchmark jump isn't.

The "should I wait?" decision tree

This is the policy I'd ship today:

def should_wait_for_opus_5(project):
    if project["needs_production_now"]:
        return "No. Benchmark Opus 4.8, Sonnet 5, and Fable 5 now."
    if project["depends_on_unconfirmed_model_id"]:
        return "Stop. Never deploy claude-opus-5 until official docs list it."
    if project["current_model_meets_sla"]:
        return "Keep the current route and make model selection configurable."
    if project["fable_quality_needed"] and project["fable_price_too_high"]:
        return "Watch Opus 5, but test current fallbacks instead of blocking launch."
    return "Build a 100-300 task eval set and wait for an official system card."

I care about six measurements after a real launch:

• Successful tasks, not benchmark headlines
• Retries per successful task
• Output tokens per success
• Tool-call errors
• Refusal or fallback behavior
• Total cost per accepted result

If Opus 5 wins those six on my workload, I migrate. If it wins a launch chart but loses cost per success, I don't.

What I'd do this week

If I ran an Opus 4.8 production service, I'd keep it running. I'd pin the exact model ID, log returned model names, and make the routing layer configurable.

If I used Sonnet 5 for most traffic, I'd continue doing that. I'd route only difficult failures to Opus 4.8 or Fable 5.

If I needed Fable-level autonomy but couldn't justify Fable pricing, I'd create the eval set now. That is the audience most likely to benefit from a future Opus 5.

If I saw an "Opus 5 benchmark" screenshot, I'd ask for the model ID, system card, harness, token budget, and reproducible endpoint. Without those, I'd treat the number as content, not evidence.

The bigger picture

Anthropic's naming is becoming more important than its version numbers. Sonnet is the scaled default. Opus is the premium enterprise and coding tier. Fable is the public long-horizon frontier. Mythos is the restricted capability tier.

Opus 5 matters only if Anthropic preserves that four-level architecture. If the company instead makes Fable the permanent successor to Opus, the question isn't "When does Opus 5 launch?" It is "Does the Opus brand still describe a long-term product?"

That is why I think the product map is a better signal than a leaked date.

If you want to switch among Anthropic and other providers through one OpenAI-compatible endpoint, that's roughly what TokenMix does. Disclosure: I work on the research side. The full source-by-source analysis is in the original Opus 5 forecast.

Bottom line

Claude Opus 5 is plausible. It is not announced. The strongest hypothesis is a $5/$25 model that sits between Sonnet 5 and Fable 5, but no date, API ID, context limit, or benchmark is ready to use as fact. I would prepare an eval and a configurable router. I would not delay a real deployment or publish invented scores.

What evidence would convince you that Opus 5 is real: an official model-catalog row, a system card, or a working API response?