Claude Opus 4.7 shipped 2026-04-16 and is the current public flagship, not a rumor
Claude Sonnet 4.8 is not announced, traced to one leaked string in Claude Code source
Claude Mythos exists and is real, but it is restricted to a handpicked group of companies
Anthropic conceded on 2026-04-16 that Opus 4.7 still trails its Mythos Preview model
Plan on Opus 4.7 and Sonnet 4.6 today, do not pause work for an undated model
The internet decided Claude has a secret super-model and a phantom Sonnet 4.8 on the way. Some of that is true. Most of it is not. I sorted the confirmed releases from the leak-blog fan fiction so you can plan a build without chasing ghosts.
What is actually shipping right now
Start with what I can point at and use today. No leaks required.
Claude Opus 4.7 went generally available on 2026-04-16. It is the current public flagship. Anthropic's own announcement frames it as a coding and long-running-task model, with a "notable improvement on Opus 4.6 in advanced software engineering" and the ability to handle "complex, long-running tasks with rigor and consistency."
The numbers Anthropic published are specific, so I trust them more than any rumor table:
CursorBench: 70 percent, versus Opus 4.6 at 58 percent.
BigLaw Bench: 90.9 percent accuracy at high effort.
Rakuten-SWE-Bench: resolves roughly 3x more production tasks than Opus 4.6.
Vision: accepts images up to 2,576 pixels on the long edge (about 3.75 megapixels), more than 3x the pixels of earlier Claude models.
Pricing for Opus 4.7 stayed put at 5 EUR-equivalent per million input tokens and 25 per million output tokens (Anthropic lists it in dollars, but the ratio is what matters for planning). That is the same as Opus 4.6, which is the rare case where a stronger model did not get more expensive. For a solo operator that detail matters: a capability jump with a flat token cost means I can upgrade my hardest workloads without re-running the math on every batch job.
On the Sonnet side, Claude Sonnet 4.6 released 2026-02-17 and is still the current Sonnet. It is the workhorse I reach for on volume tasks: cheaper, fast, good enough for most of my drafting and parsing jobs while Opus 4.7 handles the hard reasoning. I split work between the two on purpose. Routing every call to the flagship is how a small studio burns a token budget for no real gain. Sonnet 4.6 does the bulk grind, Opus 4.7 does the thinking, and that two-tier split has held steady through this whole rumor cycle.
So the lineup as of late May 2026 is simple. Opus 4.7 for the heavy work. Sonnet 4.6 for the rest. Everything past this point is either a countdown or a rumor.
The June 15 deadline that is not a rumor
Before chasing future models, clear the one date that will actually break things if you ignore it. The original Claude Opus 4 and Claude Sonnet 4 retire on 2026-06-15. After that, requests pinned to those model IDs stop working.
This is the boring, confirmed news that matters more than any leak. I have covered the specifics in the June 15 retirement of Opus 4 and Sonnet 4, so I will keep it short here.
The migration is mechanical. If you call claude-opus-4 or claude-sonnet-4 anywhere (a script, an agent, a hardcoded config, a cron job you forgot about), swap the model ID to a current one before the cutoff.
My quick checklist:
Grep your whole codebase for old model strings, not just the obvious app code.
Move heavy reasoning calls to Opus 4.7. Move volume calls to Sonnet 4.6.
Re-run your eval set after swapping. Prompts tuned for an older model sometimes drift on a newer one.
Check any third-party tools or plugins you depend on. They have their own deadlines.
The reason I lead with this instead of the shiny rumors: a hard retirement date is a fact you can act on this week. A leaked string about Sonnet 4.8 is not. One of these will cause a 2 AM outage if you ignore it. The other will not exist for months, if at all.
If your stack still references the retiring models, fix that first. Then come back and enjoy the speculation.
Claude Sonnet 4.8: where the rumor came from
Here is the headline a lot of blogs ran: Claude Sonnet 4.8 is coming. Here is the honest version: it is not announced, and the evidence is thin.
As of 2026-05-23, Anthropic has made no announcement, published no API model ID, released no benchmarks, and given no date for a Sonnet 4.8. So where did the rumor start?
A single leaked string. In late March 2026, a JavaScript source map was accidentally shipped with Claude Code, and someone spotted a "4.8" reference inside it. That string is the entire signal. Nothing came with it. No spec, no date, no confirmation.
Why I am skeptical, not dismissive
A string in a source map is not nothing. Internal builds reference unreleased things all the time. But it is also not a release. A few reasons I am keeping this in the rumor column:
The originating report flagged itself as uncertain. The post that surfaced it added a question mark, which tells you the finder was not confident either.
A prediction market on a Sonnet 4.8 release by 2026-05-24 sat at around 3 percent. The crowd does not believe it is imminent.
It would break Anthropic's versioning pattern. Opus 4.5 and Sonnet 4.5 matched. Opus 4.6 and Sonnet 4.6 matched. Jumping Sonnet to 4.8 with no 4.7 in between would skip a step Anthropic has never skipped.
That last point is the one I weigh most. Anthropic has been consistent: the Opus and Sonnet minor versions move together. A lone "4.8" that leaps past a Sonnet 4.7 is more likely a placeholder, a test ID, or an internal experiment than a near-term product. Labs leave dead strings and forward-looking IDs in their code constantly. A grep hit is a clue, not a launch date.
There is also a tell in how these posts get written. The headline says "Claude Sonnet 4.8 Release Date" with full confidence. The body, if you read it, walks it all back and admits nothing is announced. That gap between headline and body is the cleanest signal that a piece is chasing search traffic rather than reporting news.
My rule for any reader who finds this article while deciding what to build on: if a post hands you a Sonnet 4.8 benchmark table, a price, or a confirmed release date right now, it is fabricating. Treat it that way. The honest answer in late May 2026 is "no Sonnet 4.8 exists yet," and that answer will not change until Anthropic publishes a model page.
Claude Mythos and the Opus 5 question
The Mythos story is the opposite of the Sonnet 4.8 story. Mythos is real. Anthropic confirmed it. The fan fiction is in the details, not the existence.
On 2026-04-16, the same day Opus 4.7 shipped, Anthropic publicly conceded that Opus 4.7 does not match the performance of a more capable internal model called Mythos. Per Axios, the company said benchmark charts show Opus 4.7 beating ChatGPT 5.4 and Gemini 3.1 Pro, while still falling short of "its Mythos Preview model." That is an unusual move: a lab admitting in public that it is sitting on something stronger than what it just released.
Mythos is not public. Axios reports access is limited to "a handpicked group of tech and cybersecurity companies," and the model is withheld over safety concerns. Anthropic's stated goal is "a broad release of Mythos-class models" once it learns enough from deploying safeguards.
What is confirmed vs what is leak-blog guesswork
Mythos first surfaced through a large source-code and document exposure. A misconfigured data store left around 3,000 internal files reachable without authentication in late March 2026, then a researcher disclosed it and Anthropic locked it down within hours. I covered that incident in the source leak that first surfaced Mythos.
From the leak and Anthropic's own statements, here is my confidence split:
Confirmed: Mythos exists, and Anthropic calls it a major capability advance.
Confirmed: It is restricted, withheld for safety, given to a small set of companies.
Reported but unverified: Mythos is "far ahead of any other AI model in cyber capabilities," and can find and exploit vulnerabilities faster than human defenders. That phrasing comes from leaked docs, not a published benchmark.
Not confirmed: the name "Opus 5." The community attached that label. Leaked docs frame Mythos as a new capability class, not a version bump, and Anthropic has not blessed "Opus 5."
The cybersecurity angle is the most consequential part of the story and the part with the least public proof. If you want the security framing, I dug into Mythos powering Project Glasswing. Read it knowing the specifics are still reported, not benchmarked.
How I plan around all of this as a solo builder
Strip the drama and the planning answer is short. Build on what exists. Watch the rumors for free.
I run a one-person studio, so I cannot afford to architect around a model that might land in Q3, might be named something else, and might never go public in my tier. Here is the framework I actually use.
Treat the timeline as three buckets
Shipping now: Opus 4.7, Sonnet 4.6. These are the only things I let into production.
Confirmed but inaccessible: Mythos. Real, stronger, not available to me. I track it, I do not depend on it.
Rumored: Sonnet 4.8, "Opus 5," "Claude 5," and the speculated Q2 to Q3 2026 window for the next major model. None of these names are official. None get a line item in my plans.
My practical rules
Pin to current model IDs, but isolate them. I keep the model name in one place so a future swap is a one-line change, not a hunt.
Never delay a launch for an unannounced model. The cost of waiting is certain. The arrival of Sonnet 4.8 is not.
Re-run evals on every model bump. A newer model is not automatically better for your exact prompt. Measure, do not assume.
Read the leak blogs for direction, not numbers. They are useful for "Anthropic is clearly investing in cyber and long-horizon agents." They are useless for "here is the exact benchmark."
Keep a one-line note on the rumor, then move on. I track Mythos and Sonnet 4.8 in a single text file with a date and a confidence tag. That is the right amount of attention for something I cannot use yet.
The direction Anthropic is signaling is not subtle. It conceded in public that it has a stronger model than it ships. It is gating that model on safety and handing it to a narrow set of partners first. It keeps shipping minor versions of Opus and Sonnet on a tight cadence. Put those together and the realistic read is more capable models arriving steadily, with the frontier ones arriving slowly and to a few hands. None of that tells me a name or a date. All of it tells me the safe bet is to build clean on today's models and keep the swap cheap.
The signal here is that Anthropic has a stronger model than it ships, and a steady cadence of minor versions. The noise is every specific date, score, and name attached to models that do not have a public page yet. If you want the same model setup I run day to day, I keep it documented in my Claude setup blueprint.
Bottom Line
Here is the whole thing in one breath. Opus 4.7 (shipped 2026-04-16) and Sonnet 4.6 (shipped 2026-02-17) are real and usable today. Opus 4 and Sonnet 4 retire 2026-06-15, so migrate before then. Mythos is real, stronger than Opus 4.7 by Anthropic's own admission, and locked to a small set of companies for safety reasons. Claude Sonnet 4.8 is not announced and rests on a single leaked string. "Opus 5" and "Claude 5" are community labels, not confirmed names, and any next-major-model date in the Q2 to Q3 2026 range is speculation.
If you only do one thing this week, it is not waiting for a phantom model. It is grepping your codebase for the retiring model IDs and swapping them. That is the deadline with teeth.
I will keep updating as Anthropic actually announces things, not as leak blogs invent them. If you want a clean starting point that already runs on the current models and is easy to re-point when the next one lands, take a look at my Claude setup blueprint and skip the guesswork.
Top comments (0)