Claude Opus 4.7 Ships, Sonnet 4.8 Leak Raises Questions
On April 16, 2026, Anthropic released Claude Opus 4.7 (model ID: claude-opus-4-7) to general availability across its own API as well as Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The pricing held steady at $5 per million input tokens and $15 per million output tokens — identical to Opus 4.6 — but the request-cost multiplier jumped to 15× after the promotional window closed on April 30. What makes this release unusual is not what shipped, but what did not: a matching Sonnet model. Instead, a string buried in a 512K-line TypeScript source map suggests Anthropic may be skipping straight to Sonnet 4.8. Here is what is confirmed, what is speculative, and why it matters.
What Opus 4.7 Actually Improves
Anthropic positioned Opus 4.7 as an iterative-but-meaningful upgrade focused on three areas that matter to production engineers: hard software-engineering tasks, high-resolution vision, and long-running agentic workflows with built-in self-verification.
On a 93-task internal coding benchmark that stresses multi-file refactoring, test generation, and debugging of real open-source repositories, Opus 4.7 posted a 13% lift over Opus 4.6. Four of those tasks were solved by Opus 4.7 but by neither Opus 4.6 nor Sonnet 4.6, suggesting the model crossed a qualitative threshold on certain complex reasoning chains rather than simply getting faster at easy problems.
Vision improvements target higher-resolution inputs — useful for document extraction, architectural diagram parsing, and satellite imagery analysis where previous models would downsample and lose critical detail. The long-context agentic benchmark tells a similar story: Opus 4.7 scored 0.715 overall, with its General Finance module hitting 0.813 compared to Opus 4.6’s 0.767. For pipelines that run Claude autonomously for dozens of steps — codebase migrations, security audits, multi-hour research tasks — this is the metric that justifies the 15× billing multiplier.
Self-verification is the quiet feature here. Opus 4.7 can check its own intermediate outputs within a single generation, reducing hallucination cascades in multi-step tool-use chains. If you have ever watched an agent confidently build on a wrong assumption for six consecutive tool calls, you understand why this matters.
Opus 4.6 vs. Opus 4.7 — Benchmarks at a Glance
| Metric | Opus 4.6 | Opus 4.7 |
|---|---|---|
| 93-task coding benchmark | baseline (100%) | +13% lift, 4 novel solves |
| Research agent overall | 0.670 | 0.715 |
| Research agent long-context | — | Best-in-class (undisclosed) |
| General Finance module | 0.767 | 0.813 |
| Input / output pricing | $5 / $15 per MTok | $5 / $15 per MTok |
| Request multiplier (post-promo) | 10x | 15x |
| Vision resolution | Standard | Higher native resolution |
The Glasswing Safeguard Layer
Internally, the Opus 4.7 project carried the codename Glasswing — a reference not to performance but to the cybersecurity safeguards Anthropic layered into the model before release. Glasswing encompasses a set of eval-driven refusal thresholds tuned specifically for cyber-offense capabilities: generating exploit payloads, crafting socially-engineered phishing at scale, and automating vulnerability discovery in target systems.
Anthropic has been increasingly vocal about the dual-use risk of capable coding models. Glasswing appears to be the engineering response: the model can write security tooling and patch code, but it degrades noticeably when prompted to generate offensive toolchains. How well this boundary holds under adversarial prompting is an open question — and one that every team deploying Opus 4.7 in a security-adjacent pipeline should red-team themselves before trusting in production.
The Sonnet 4.8 Leak — What We Know
On March 31, 2026 — two weeks before Opus 4.7 launched — a developer inspecting the source map of Claude Code npm package v2.1.88 found the string sonnet-4-8 in a “forbidden strings” list. That list also contained opus-4-7 (which subsequently shipped) and mythos-preview (which has not). The source map contained approximately 512,000 lines of TypeScript, and the forbidden-strings mechanism appeared to be a client-side filter preventing users from requesting unreleased models through Claude Code.
As of May 28, 2026, Sonnet 4.8 has not been released. There is no model card, no API model ID, no benchmark data, and no official acknowledgment from Anthropic. A Polymarket contract titled “Sonnet 4.8 by May 24” closed at 3%, meaning the market essentially treated it as vapor for that window.
Three readings are circulating:
- Internal codename: Engineering teams routinely add identifiers for models that are months away. The forbidden-string list is a safety net, not a release calendar.
- Skipped Sonnet 4.7: Anthropic has always paired Opus and Sonnet at the same version (4.5/4.5, 4.6/4.6). Opus 4.7 has no matching Sonnet 4.7 — an unprecedented gap. One hypothesis is that Sonnet 4.7 was absorbed into a larger training run and will ship as 4.8 with a bigger capability delta.
- Imminent release: The presence of the string in production client code, alongside an already-shipped model, suggests the model exists behind Anthropic’s API gateway and could be activated at any time.
My Take: Sonnet 4.8 Is Real and Likely Imminent
The skipped-sonnet hypothesis is the most plausible explanation. Anthropic’s version-pairing convention exists for a reason — it simplifies pricing tiers, documentation, and customer communication. Breaking that convention is not a decision taken lightly. The fact that no Sonnet 4.7 appeared alongside Opus 4.7 strongly suggests the Sonnet branch was rebased onto a later training checkpoint and will ship with enough improvements to warrant a new minor version number.
The Polymarket closure at 3% is noise, not signal. Prediction markets are thin for niche AI model releases, and a two-week window was always unlikely. The real indicator is structural: Anthropic needs a mid-tier model to fill the pricing gap between Haiku and Opus. Without a new Sonnet, the product lineup has a capability hole that competitors (GPT-5-class models, Gemini 2.7 Pro) will exploit aggressively.
As for mythos-preview, that string likely refers to Anthropic’s most powerful internal model — one with limited release due to cyber risk concerns. If Mythos is real, it sits above Opus in the hierarchy, which further explains why Anthropic might renumber Sonnet: the product line is being reshuffled to make room for a new top tier.
If you are building production systems on the Claude family, plan your architecture around model ID flexibility. Hard-coding claude-opus-4-7 into your pipeline is fine today, but expect a claude-sonnet-4-8 (or whatever it lands as) within weeks. The request-cost multiplier — likely 5× to 8×, based on historical Sonnet pricing — will make it the default choice for most workloads that do not require Opus-tier reasoning depth.
What This Means for Cloud and AI Engineers
If you are already on Opus 4.6, the upgrade path to 4.7 is straightforward — same model ID prefix, same API surface, same pricing at the token level. The gotcha is the request multiplier: if your workload generates many small requests (agentic loops, parallel tool calls, streaming chat), your effective bill will be higher than the per-token numbers suggest. Run a cost comparison on your actual request distribution before flipping the default.
The self-verification feature is worth testing on any multi-step agent pipeline you run in production. Early reports suggest it reduces error propagation by 15-25% on complex tool-use chains, but the verification steps themselves consume tokens. Net cost impact will vary by workload.
For teams evaluating whether to wait for Sonnet 4.8: if your use case tolerates Opus 4.6’s reasoning depth, you are not missing much by staying put. If you need Opus 4.7’s improvements on hard coding tasks or long-context research, upgrade now. Sonnet 4.8 — whenever it ships — will not match Opus 4.7 on the hardest tasks. It never does. That is the entire point of the tier split.
FAQ
- Is Claude Opus 4.7 available on all cloud platforms? Yes. It launched on the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry simultaneously on April 16, 2026. All four platforms use the same model ID:
claude-opus-4-7. - What does the 15× request multiplier mean in practice? Each API call to Opus 4.7 counts as 15 requests against your rate limits, compared to 10× for Opus 4.6. If your tier allows 1,000 requests per minute, you can make roughly 66 Opus 4.7 calls per minute instead of 100 Opus 4.6 calls. Token pricing per million tokens is unchanged.
- Is Sonnet 4.8 confirmed by Anthropic? No. As of May 28, 2026, Anthropic has made no public statement about Sonnet 4.8. The only evidence is a leaked string in Claude Code’s source map. Treat it as high-probability speculation, not fact.
- What is Mythos Preview? A model identifier found alongside
sonnet-4-8in the Claude Code forbidden-strings list. It is described as Anthropic’s most powerful model but with restricted availability due to cybersecurity concerns. No further details have been confirmed. - Should I migrate from Opus 4.6 to 4.7 immediately? Only if your workload depends on the specific improvements — hard coding tasks, long-context agentic performance, or high-resolution vision. The token pricing is identical, but the higher request multiplier means bursty workloads will hit rate limits sooner. Test with a traffic shadow before a full cutover.