Claude Sonnet 4.8: Release Status & What's Rumored (2026)

If you searched "Claude Sonnet 4.8" you likely saw a single low-confidence post pointing at a leaked string in Claude Code's source. That is the entire basis of the rumor. There is no Anthropic blog post, no model id, no spec sheet, and no date. This page is the honest status check, refreshed on May 25, 2026: what Anthropic is actually shipping in the Sonnet tier, exactly where the 4.8 reference came from and why it is weak, what (if anything) the pattern suggests, what teams should do today, and how Anthropic's cadence stacks up against OpenAI, Google, and DeepSeek's recent releases. We rewrite this page in place the moment a real Sonnet 4.8 lands, so it stays the current answer rather than a dated guess.

Is Claude Sonnet 4.8 out?

No. As of May 25, 2026, Claude Sonnet 4.8 still does not exist in any form a developer can use:

• No announcement. The most recent posts on anthropic.com/news (May 22: Project Glasswing update; May 19: KPMG partnership; May 18: Stainless acquisition; May 13: Claude for Small Business; May 6: usage-limit increase and SpaceX compute deal) include nothing about a new Sonnet. The most recent Sonnet announcement is still Claude Sonnet 4.6 from February 17, 2026.
• No API model id. The official Anthropic model list still documents claude-sonnet-4-6 as the current Sonnet and claude-opus-4-7 as the flagship. There is no claude-sonnet-4-8 id, alias, or snapshot. There is no claude-sonnet-4-7 either — a point worth pausing on (see below).
• No benchmarks, no specs. No context window, pricing, knowledge cutoff, or capability list — none of it has been published, because Anthropic has published nothing.
• The market agrees. The Manifold prediction market "Will Claude Sonnet 4.8 be released before May 24?" closed at 3%. That's a clean, public, money-on-the-line signal that traders did not expect Sonnet 4.8 in that window — and it resolved NO on schedule.

Anyone presenting Sonnet 4.8 specifications, a benchmark table, a price, or a confirmed release date today is fabricating. We are not going to do that. When real numbers exist, this section becomes the release record.

What Claude models are actually shipping right now?

The models you should actually care about today are Claude Sonnet 4.6 for the balanced speed/intelligence tier and Claude Opus 4.7 for frontier reasoning and agentic coding. Both are generally available now. The current lineup, with verified pricing from Anthropic's own model list (May 25, 2026):

Sonnet 4.6 (released February 17, 2026) is positioned by Anthropic as delivering performance that previously required reaching for an Opus-class model on many real-world tasks, while Opus 4.7 (released April 16, 2026) remains the recommendation for the most demanding reasoning. For coding-agent and tool-use work specifically, Sonnet 4.6 is the default workhorse and Opus 4.7 is the escalation path. That is the practical choice today; Sonnet 4.8 is not on the menu.

Note: a separate research-preview model called Claude Mythos Preview exists for Anthropic's Project Glasswing defensive-cybersecurity program. It launched April 7, 2026, is invitation-only, priced at $25 / $125 per MTok (reported, gated to Glasswing partners), and is not a Sonnet successor — don't confuse it with the 4.8 rumor. We come back to Mythos below because it's the more meaningful "next-gen" signal from Anthropic this quarter.

Where did the Sonnet 4.8 rumor come from?

One thread, and a weak one:

1. A leaked string in Claude Code's source. Someone reported spotting a "Sonnet 4.8" reference inside the Claude Code client's source bundle on March 31, 2026 — the day a source-map file accidentally shipped with the @anthropic-ai/claude-code npm package and exposed roughly 512,000 lines of TypeScript including references to unreleased models and internal features. The "Sonnet 4.8" string sits inside that dump alongside other internal references. That is the whole signal. There is no accompanying announcement, model card, API id, or date.
2. The originating post flagged itself uncertain. The tweet that surfaced it carried an explicit question-mark qualifier — the lowest corroboration of the recent model-leak batch. It was framed as "unconfirmed," not reported as fact.
3. The cadence assumption. Anthropic has shipped Sonnet point releases at a steady pace (4.5, then 4.6). People reasonably expect another Sonnet release eventually. That is anticipation, not a leak — and "eventually" is not a date.

A model name appearing in a client's strings can mean a planned release, an internal placeholder, a routing experiment, a build artifact, or a name that gets changed before launch. On its own it is a clue, not a product.

Why skipping from 4.6 to 4.8 would break Anthropic's pattern

Here is the part most "Sonnet 4.8" coverage misses, and the strongest reason to be skeptical of an imminent launch: Anthropic has never shipped a Sonnet that is two minor versions ahead of the latest Opus. Sonnet has consistently been at the same minor version as Opus, or one behind:

• Opus 4.5 / Sonnet 4.5 — same minor version.
• Opus 4.6 / Sonnet 4.6 — same minor version.
• Opus 4.7 (current flagship, April 16, 2026) / no Sonnet 4.7 has shipped.

For Sonnet to jump straight to 4.8 without a 4.7 existing would be a first. The simpler, more pattern-consistent expectation is that the next Sonnet release matches the current Opus version — i.e. Sonnet 4.7 — and that the leaked string is either an internal placeholder, a planning artifact, or a name that gets corrected before launch. This is exactly the argument the lead Manifold trader laid out, and it is why the market resolved at 3% for "Sonnet 4.8 by May 24," not 30%.

None of this rules out a Sonnet 4.8 eventually. It just makes "any day now, here's the spec sheet" coverage smell like SEO bait rather than real news.

What does Anthropic's 4.x release cadence tell us about timing?

The cadence is the most useful evidence we have. Here is the verified Claude 4.x release record, with the gaps between releases:

Two things to read off this table. First, the prior Sonnet-after-Opus gap was 12 days (Opus 4.6 → Sonnet 4.6). It is now 39 days since Opus 4.7 with no Sonnet successor in sight, so whatever the next Sonnet is — 4.7 or 4.8 — Anthropic has already broken from the tight cascade they ran in February. Second, Sonnet 4.5 to Sonnet 4.6 took ~141 days. The 141-day mark from Sonnet 4.6 (Feb 17) lands on July 8, 2026. If history repeats, that is a more grounded expectation window than "before May 24."

The corollary: Anthropic is not late by their own track record, they are mid-window. Aggregator pages claiming Sonnet 4.8 was "expected early-to-mid May" were extrapolating from the Opus-to-Sonnet-4.6 cascade (12 days) and ignoring the slower Sonnet-to-Sonnet cadence. The two numbers tell different stories. Plan on the longer one.

How does Anthropic's roadmap compare to GPT-5.5, Gemini, and DeepSeek V4?

Anthropic's quiet Sonnet front in May looks louder when you set it next to what the rest of the frontier shipped in the same window. Three competitors moved in April:

• OpenAI shipped GPT-5.5 on April 23, 2026 across ChatGPT Plus, Pro, Business, and Enterprise, with GPT-5.5 Pro available in the API. OpenAI's announcement positions it as their "smartest and most intuitive" model, with the headline gains in agentic coding, knowledge work, and research — exactly the workload Sonnet has owned.
• Google shipped Gemini 3.1 Pro on February 19, 2026 and expanded availability across AI Pro and Ultra subscribers on May 4, 2026, broadening developer access across its tiering rather than pushing a brand-new flagship.
• DeepSeek shipped V4 Preview on April 24, 2026 with two open-weights variants (V4-Pro at 1.6T total / 49B active, V4-Flash at 284B / 13B), both with 1M-token context and dual Thinking / Non-Thinking modes under MIT license — a direct shot at closed-source agentic-coding workloads at a fraction of the price.

So in roughly a one-week window in late April, the agentic-coding tier got a major OpenAI flagship and an open-weights frontier release that anyone can self-host. Anthropic's response has been Opus 4.7 (April 16) and Mythos Preview (April 7, gated to Project Glasswing), with the Sonnet tier untouched since February. That competitive pressure is the strongest non-leak argument for a Sonnet update soon — and the simultaneous reason Anthropic might prefer to ship something materially better rather than rushed.

Practical reading for teams: the agentic-coding model market has a real chooser's dilemma right now in a way it didn't six months ago. Sonnet 4.6 remains a strong default, but GPT-5.5 and DeepSeek V4 are both viable on the same workloads. If you're benchmarking, run all three.

What other signals from Anthropic actually matter?

If you're trying to read the tea leaves more carefully than "one leaked string," here is the better-corroborated picture as of May 25:

• Claude Mythos Preview (April 7, 2026) is the substantive "next-gen" announcement. The model ships under Project Glasswing to ~12 founding partners and ~40 critical-infrastructure operators, priced at $25 / $125 per MTok (reported, gated to Glasswing partners). Anthropic identified "thousands of zero-day vulnerabilities" with it and committed $100M in usage credits. Mythos is what an Anthropic capability push looks like in 2026 — not a v4.8 number bump, but a new naming line that breaks from the Opus/Sonnet/Haiku tier system. Watch for whether Mythos capabilities cascade down into the Sonnet tier in the next release.
• Claude Code model defaults (effective April 23, 2026) tell you what Anthropic actually wants developers using right now. On the Anthropic API and AWS, the opus alias resolves to Opus 4.7 and sonnet resolves to Sonnet 4.6. Enterprise pay-as-you-go and API users now default to Opus 4.7; Max and Team Premium default to Opus 4.6; Pro and Team Standard default to Sonnet 4.6. Claude Code also auto-falls back from Opus to Sonnet at usage thresholds for certain Max users — the routing behaviour is built around the existing two-model lineup, not a placeholder for an unreleased Sonnet 4.8.
• Anthropic's news feed in May 2026 is dominated by enterprise partnerships (KPMG, PwC, Gates Foundation, Blackstone joint venture), infrastructure (SpaceX compute deal, higher usage limits), and platform/product moves (Claude for Small Business, the Stainless acquisition). No dev-rel hints about Sonnet timing have appeared on the official channels. Where you see "Sonnet 4.8 expected this week" claims, they are extrapolations from the source-map leak, not corroborated signals.
• What is not visible matters too: no system card, no waitlist page, no quiet API surface where claude-sonnet-4-7 or claude-sonnet-4-8 returns anything other than a 404, no Anthropic engineer posting "we're cooking" hints on social. Real Anthropic releases tend to come with at least a few of those leading indicators. None are present.

Put together, the most honest summary is: Anthropic is not pre-announcing a Sonnet update, and there's no operational evidence that one is days away. There is one weak signal (the source-map string) and one structural reason to expect something in the Sonnet tier eventually (competitive pressure plus the broken Opus-to-Sonnet cadence). That's it.

What might the next Sonnet bring if the pattern holds?

This is the only forward-looking section, and it is explicitly an extrapolation from Anthropic's public Sonnet 4.x cadence — not specifications, and deliberately number-free. Given how weak the underlying signal is, treat this with even more caution than usual. If the established Sonnet pattern continues, the next point release would plausibly iterate on the same axes Anthropic has emphasized release to release: agentic and coding reliability, tool-use and long-horizon task handling, and the speed/cost balance that defines the Sonnet tier. None of that is promised, none of it is dated, and we are publishing no benchmark numbers, no context-window figure, and no price — because inventing them would be the exact thing this article exists to counter. Assume nothing for production planning until Anthropic ships.

Sonnet 4.8 vs Sonnet 4.6 / Opus 4.7 — which should I use?

Use Sonnet 4.6 for balanced work, or Opus 4.7 for frontier work. There is no Sonnet 4.8 to compare against — any benchmark, price, or capability comparison involving "Sonnet 4.8" today is fabricated, because the model does not exist. Concretely: reach for claude-sonnet-4-6 as the default for coding agents, tool use, and high-throughput tasks; escalate to claude-opus-4-7 when a task demands the deepest reasoning, large-scale refactoring, or coordinating multiple agents. If a real Sonnet 4.8 (or 4.7) ships and is materially better, migrating is a model-id swap plus a re-eval — cheap compared to stalling on a model that does not exist.

Should you wait for Sonnet 4.8 or ship on Sonnet 4.6 now?

Ship on Sonnet 4.6 now. "Wait for the next version" is almost always the wrong call when (a) there is no announced version, let alone a date, (b) the prediction market resolved at 3% for the most recent dated window, (c) the current model is strong and generally available, and (d) the alternative is shipping nothing in the meantime. Sonnet 4.6 is a current, capable model serving production traffic today, and Opus 4.7 is right there when you need more headroom. Sonnet's audience is overwhelmingly people building coding agents and tool-using assistants — if that is you, the right move is to build on what exists and keep your model-selection layer swappable, not to pause a roadmap on a single uncorroborated string. For the broader picture of how Claude fits into agent stacks alongside the other frontier models, see our AI coding agents complete guide for 2026, and for a direct head-to-head with the competitive set on coding tasks, our Grok 4 vs Claude Opus 4.7 vs Gemini 2.5 Pro on coding comparison. If you're stress-testing Opus 4.7 with task budgets in particular, our writeup on Opus 4.7 task budgets for agentic coding goes deep on what we've learned in production.

What does waiting actually cost you if Sonnet 4.7 slips by 60 or 180 days?

"Wait for the next Sonnet" sounds free. It isn't. A "wait" decision usually shows up as three slowdowns: paused work on agent prompt templates, eval suites, and tool-use scaffolds you'd want to re-test against the new model; deferred API-spend bumps stuck in a second budget cycle; and competitive ground given to teams who shipped — every week, your competitors on Sonnet 4.6 / Opus 4.7 / GPT-5.5 / DeepSeek V4 are accumulating production eval traces and edge-case prompts you don't have yet.

If the next Sonnet ships in 60 days (mid-late July, the longer historical-cadence reading), the "I waited" team is two engineer-months behind and still owes the migration work. If it slips 180 days, the team that built on Sonnet 4.6 likely doesn't care anymore — they're either on Opus 4.7 for hard tasks or already evaluating GPT-5.5 and DeepSeek V4 as alternates.

Concrete recommendation for this week, regardless of when Sonnet 4.7/4.8 lands: build on Sonnet 4.6 with Opus 4.7 as the documented escalation path; set ANTHROPIC_MODEL or the model field via server-managed settings so a future swap is one config change; keep your model layer swappable so you can re-run the same evals against GPT-5.5 and DeepSeek V4-Flash; and have ~50-200 representative eval tasks ready so a new model can be re-ranked inside a week. The day Anthropic ships is not the day to start writing evals.

How will you know when Sonnet 4.8 is real?

Watch the two primary sources, not aggregators, SEO posts, or screenshots of source strings:

• anthropic.com/news — the official announcement, with positioning and benchmarks, appears here first.
• The Anthropic API model list — a claude-sonnet-4-8 id (or claude-sonnet-4-7, more likely first) with documented pricing, context window, and cutoff is the definitive, build-on-it signal.

When a real next-generation Sonnet lands, this article gets rewritten in place — same URL — into the full release guide: actual model id, verified benchmarks, pricing, context window, and migration notes from 4.6. Bookmarking this page means you get the accurate version the day it matters, instead of the speculation circulating now.

If your team is moving fast on the Claude stack — evaluating each Sonnet and Opus release, running your own benchmarks, and keeping coding-agent infrastructure current — that is real engineering work. Codersera matches you with vetted remote developers experienced building on Claude and other frontier models in production, with a risk-free trial so you can validate technical fit before committing.

FAQ

Is Claude Sonnet 4.8 out?

No. As of May 25, 2026 there is no Claude Sonnet 4.8 announcement, no API model id, no benchmarks, and no release date. The current Sonnet model is Claude Sonnet 4.6 (claude-sonnet-4-6, $3 / $15 per MTok, 1M context) and the current flagship is Claude Opus 4.7 (claude-opus-4-7, $5 / $25, 1M context). The Manifold market for a release before May 24 resolved at 3%.

Why is there no Sonnet 4.7? Why would Anthropic jump straight to 4.8?

They probably wouldn't. Anthropic has consistently kept Sonnet at the same minor version as Opus or one behind (Opus 4.5 / Sonnet 4.5, Opus 4.6 / Sonnet 4.6). Skipping from Sonnet 4.6 straight to Sonnet 4.8 without a 4.7 would be a first. The simpler explanation for the leaked "4.8" string is that it is an internal placeholder, planning artifact, or name that gets corrected before launch — and that the next Sonnet release is more likely to be 4.7, matching the current Opus version.

How long does Anthropic usually take between Sonnet releases?

About 4-5 months in the 4.x family. Claude Sonnet 4.5 shipped September 29, 2025 and Claude Sonnet 4.6 shipped February 17, 2026 — a ~141-day gap. The Opus-to-Sonnet cascade within a single minor version has been much faster (12 days between Opus 4.6 and Sonnet 4.6), but the Sonnet-to-Sonnet gap is the more relevant historical anchor. From Sonnet 4.6's release that puts the next Sonnet point release plausibly in early-to-mid July 2026 if Anthropic stays on the same cadence — but they are under no obligation to, and have not published a timeline.

How does Anthropic's roadmap compare to OpenAI, Google, and DeepSeek this spring?

Anthropic shipped Opus 4.7 on April 16 and Claude Mythos Preview (Project Glasswing only) on April 7. OpenAI shipped GPT-5.5 across ChatGPT Plus/Pro/Business/Enterprise on April 23. Google shipped Gemini 3.1 Pro on February 19 and broadened access on May 4. DeepSeek shipped V4 Preview (V4-Pro and V4-Flash, both 1M context, MIT-licensed open weights) on April 24. Anthropic's Sonnet tier is the one frontier-tier slot that did not get an update in the late-April window, which is part of why the Sonnet 4.8 rumor has had legs even without official corroboration.

Should I wait for the next Sonnet or build now?

Build now on Sonnet 4.6 (with Opus 4.7 as the escalation path), and keep your model-selection layer swappable. If the next Sonnet ships in 60 days, the cost of having waited is roughly two engineer-months of work-not-done plus a missed window to gather production eval data; if it slips 180 days, the cost compounds. The migration when it ships is typically a model-id swap and a re-run of your eval suite — small compared to pausing your roadmap.

What is Claude Mythos, and is it related to Sonnet 4.8?

Claude Mythos Preview is Anthropic's invitation-only model launched April 7, 2026 alongside Project Glasswing, the defensive-cybersecurity initiative with ~12 founding partners (AWS, Google, Microsoft, NVIDIA, JPMorganChase, Apple, Cisco, CrowdStrike, Palo Alto Networks, the Linux Foundation, Broadcom, and Anthropic). Pricing is $25 / $125 per MTok (reported, gated to Glasswing partners) and access runs through Project Glasswing. Mythos is structurally distinct from the Sonnet/Opus/Haiku tier system — it's not a Sonnet successor, but it is the most meaningful next-gen signal from Anthropic this quarter.

Sonnet 4.8 vs Sonnet 4.6 / Opus 4.7 — which should I use?

Use Sonnet 4.6 or Opus 4.7 — Sonnet 4.8 does not exist, so any comparison involving it today is fabricated. Pick claude-sonnet-4-6 ($3 / $15 per MTok, 1M context) for balanced speed/intelligence and most coding-agent work; escalate to claude-opus-4-7 ($5 / $25, 1M context) for the deepest reasoning and large-scale refactoring.

When will Claude Sonnet 4.8 be released?

There is no announced date. The Manifold prediction market for a release before May 24, 2026 resolved at 3%. Anthropic's steady Sonnet cadence (4.5 to 4.6 was ~141 days) makes a future point release plausible — most likely 4.7 first, to match the current Opus version — but Anthropic has published no timeline. Any specific "Sonnet 4.8 release date" circulating now is speculation.

Is the Sonnet 4.8 leak credible?

It is weak. The only basis is a reported "Sonnet 4.8" string in Claude Code's source (the March 31, 2026 source-map leak), and the originating post flagged itself as uncertain. Combined with the pattern-break it would represent (no Sonnet 4.7 has shipped), it's better treated as anticipated, not announced.

What is the Claude Sonnet 4.8 API model id?

There isn't one. The Anthropic model list documents claude-sonnet-4-6 as the current Sonnet and claude-opus-4-7 as the flagship. No claude-sonnet-4-8 or claude-sonnet-4-7 id, alias, or snapshot exists. A documented id on the official model list is the signal that it has actually shipped.

Is Sonnet 4.8 better than Sonnet 4.6 or Opus 4.7?

This question cannot be answered — there are no Sonnet 4.8 benchmarks because there is no Sonnet 4.8. Any claimed performance comparison today is invented. We will publish a real comparison here only when verifiable numbers exist.

Where is the official Sonnet 4.8 announcement?

There isn't one. The authoritative places it would appear are anthropic.com/news (the announcement) and the Anthropic API model list (the model id and specs). Until something is published there, treat Sonnet 4.8 as a low-corroboration rumor, not a product.

Is Claude Mythos Preview the same as Sonnet 4.8?

No. Claude Mythos Preview is a separate research-preview model gated to Project Glasswing partners for defensive-cybersecurity work, priced at $25 / $125 per MTok and invitation-only. It is not a Sonnet successor and is not relevant to general agentic-coding workloads.