Issue #00  ·  Volume I  ·  The Founding Issue
A dry-run — three weeks to stress-test the format at maximum size.
2026 — 04 — 16
Period 2026-03-27 → 2026-04-16
56 Releases · 21 Days · 1 Shadow
01
Shipped.  ·  Volume I  ·  The Founding Issue  ·  00 Shipped.

The Shadow Release

Three weeks of Anthropic, in one read. The model they shipped. The model you can't have. And the distance between them — the new normal.

The Lead Opus 4.7 vs. the chart
The Investigation Glasswing
The Survey Claude Code, v26
The Term shadow release
p.02 Shipped. — 00 — Apr 2026
Front of book
02
In this issue,
6 Front-of-book pieces
56 Release-log entries
01
The Open A thing Anthropic does now that doesn't have a name A flagship, and a better one you can't have — in the same breath.
p.03 02
By the Numbers The shape of three weeks 56 releases. 26 versions. $100M in credits. The frontier in ten digits.
p.04 03
The Lead Story The chart that wasn't about Opus 4.7 The fourth bar. Taller than every released model. Named ‘Mythos Preview.’ Gated.
p.06 04
Investigation Glasswing Twelve companies. $100M in credits. A 27-year-old OpenBSD bug. And a precedent.
p.09 05
Feature The agent stack got serious Managed Agents. The advisor tool. The ant CLI. Your loop just became a glue layer.
p.12 06
Timeline 21 days, 56 releases Read the bumps. The pattern of a lab that does not rest.
p.15 07
Survey Claude Code grew a body Twenty-six versions. Routines. /ultrareview. The CLI is becoming the work environment.
p.17 08
Term of the Issue shadow release A word for the thing Anthropic does now. It needed one.
p.19 09
Quiet on the Wire Absences worth watching No Haiku 5. No Opus 4.7 fast-mode. No date for Mythos. Four deprecation clocks.
p.20 10
The Release Log A 1:1 mirror of every Anthropic release Seven sections. Fifty-six entries. The reference engineers grep on Tuesday.
p.21
p.03 Shipped. — 00 — Apr 2026
The Open

There's a thing Anthropic does now that doesn't have a name. They release a flagship model, and in the same breath, they tell you about a better one you can't have.

Opus 4.6 in November. Opus 4.7 today. Mythos in between, gated, available only to a handful of security partners.

Call it the shadow release. The model that ships sets the floor. The model that doesn't reveals the ceiling. The space between them is where everyone who builds with this stuff now lives.

Three weeks of velocity below. Fifty-six releases — models, agents, infrastructure, a security pact with twelve names you'd recognize, twenty-six versions of Claude Code, a tokenizer change you'll feel in your bill. Plus the shadow.

p.04 Shipped. — 00 — Apr 2026
By the numbers

The shape of three weeks.

Releases shipped 56 March 27 → April 16. One lab. No rest day above two.
Claude Code versions 26 v2.1.85 through v2.1.111. The CLI is becoming an IDE.
Agent SDK releases 20 Ten Python. Ten TypeScript. Parity maintained.
Frontier held back 1 Mythos Preview. Gated. The ceiling.
Cybersecurity benchmark — cybergym 83.1% Mythos Preview on CyberGym vulnerability reproduction. Opus 4.6, the previous frontier, sits at 66.6%. The gap is not marginal.
Model credits pledged — glasswing $100M To the twelve consortium members. Plus $4M to OpenSSF, Apache, and Alpha-Omega — the open-source maintainers who don't have security teams.
Oldest bug Mythos found 27 yr A memory-corruption flaw in OpenBSD. Survived decades of human review.
Firefox shell exploits 181 Produced by Mythos in testing. Opus 4.6 produced 2 in the same trial.
Tokenizer change 1.35× Same input may consume up to 35% more tokens on Opus 4.7. Re-tune your prompts. The bill will tell you.
Opus 4.7 pricing (unchanged) $5/$25 Per million input / output tokens. Same as 4.6 — but the tokenizer is not.
Sonnet 4 / Opus 4 retire 60 days June 15, 2026. If you haven't migrated, the clock started two days ago.
Glasswing public report 90 days Until the consortium's first progress disclosure. Read it carefully — the methodology will be precedent.
The Lead Story 01
p.06  —  Report by Eddie Belaval

The chart that wasn't about Opus 4.7.

Opus 4.7 ships. Same price. Better vision. Sharper coding. And then, on the fourth bar of the launch chart, a model you cannot buy — taller than everything else.

Report  ·  Opus 4.7 launch  ·  Apr 16, 2026
Floor & ceiling  ·  2,576 px vision  ·  xhigh default  ·  1.0 – 1.35× tokenizer
↗   Apr 16   ·   GA   ·   Opus 4.7

Opus 4.7 went generally available this morning. Same pricing as 4.6 — five dollars per million input, twenty-five per million output. Better at coding, sharper on hard software engineering, with the kind of self-verification that previously required you to paper over with scaffolding. Vision tripled in resolution: 2,576 pixels on the long edge, more than three times what prior versions could see. The model can now read dense screenshots and pixel-perfect references without you down-sampling first.

A new effort level called xhigh slotted in between high and max. Claude Code now defaults to xhigh on every plan. If you're paying attention to your bills, that change matters as much as anything else in the launch — xhigh produces more thinking tokens than high, and Anthropic raised the default before most users will notice.

The migration notes mention a tokenizer change: the same input may consume 1.0 to 1.35× more tokens depending on what you feed it. Re-tune your prompts. The bill will tell you which way it broke.

The vision change deserves its own paragraph. At 2,576 pixels on the long edge — roughly 3.75 megapixels — Opus 4.7 can read a full-density desktop screenshot without you cropping or scaling. Diagrams that used to need OCR pre-processing can now go in raw. UI mockups can be referenced pixel-for-pixel for code generation. This isn't a knob; it's a workflow change for anyone who's been hand-feeding the model lower-resolution slices.

But the headline isn't the model.

The headline is the chart Anthropic published next to the model.

The fifth bar, the Mythos bar.

SWE-bench Verified · production software engineering
Higher is better · published by Anthropic, Apr 16
Gemini 3.1 Pro
80.6%
Claude Opus 4.6
80.8%
GPT-5.4
84.5%
Claude Opus 4.7 Today
87.6%
Mythos Preview Frontier · Gated
93.9%
Fig. 01 The orange bar is the frontier you can read about but cannot call. Opus 4.7 is the floor of what you can buy this week. Mythos is the ceiling of what exists. The 6.3-point gap is the new normal — and Anthropic has decided, for now, that the distance is a feature.
↗   Lead   ·   Continued

It shows Opus 4.7 beating Opus 4.6. Beating GPT-5.4. Beating Gemini 3.1 Pro across the relevant benchmarks. And then it shows Opus 4.7 losing to a fourth bar labeled Mythos Preview. The fourth bar is taller than every other bar. The fourth bar is the model you can't use.

The frontier is in a vault, working on the OSS-Fuzz corpus, finding 27-year-old vulnerabilities in OpenBSD. Opus 4.7 is the floor of what you can buy this week. Mythos is the ceiling of what exists. The distance between them is the new normal — and Anthropic has decided that the distance, for now, is a feature, not a bug.

The implication is institutional, not technical. For the first time, the lab has an internal precedent for "we found something too dangerous to ship in the standard channel." That precedent will be invoked again. Probably soon. The next time it happens, the chart will look familiar.

Sidebar  ·  A profile of the model you can't use

The Mythos Bar, in detail.

Where Opus 4.6 hits 66.6% on the CyberGym vulnerability-reproduction benchmark, Mythos Preview hits 83.1%. Where Opus 4.6 produced two working Firefox shell exploits in several hundred attempts, Mythos produced 181. On the OSS-Fuzz corpus, Mythos achieved full control-flow hijack on ten separate, fully patched targets.

The vulnerabilities Mythos has surfaced are subtle, old, or both:

You can't have it. Mythos is gated to twelve security partners under Project Glasswing. Anthropic's stated reason is risk: a model that finds zero-days in OpenBSD that survived twenty-seven years of human review is a model you don't put on the open API. Whether that calculus holds — for how long, and under what pressure — is the open question of the year.

✦   ✦   ✦
Companion to the Lead — Fig. 03 Where each model wins
Cross-benchmark sweep

The pattern is the news.

Eight benchmarks Anthropic published in the Opus 4.7 launch, with comparable scores from competitor models where reported. Bold = winner. Orange bold = Mythos Preview leads.

Benchmark Opus 4.6 Opus 4.7 GPT-5.4 Gemini 3.1 Pro Mythos
SWE-bench Verified // general SWE 80.8% 87.6% 84.5% 80.6% 93.9%
SWE-bench Pro // agentic SWE — 64.3% 57.7% 54.2% —
GPQA Diamond // graduate reasoning 91.3% 94.2% 94.4% 94.3% 94.5%
MCP-Atlas // scaled tool use 75.8% 77.3% 68.1% 73.9% —
CursorBench // IDE agentic 58.0% 70.0% +12 — — —
XBOW Visual-Acuity // vision 54.5% 98.5% +44 — — —
BigLaw Bench // Harvey, legal — 90.9% — — —
CyberGym // vulnerability repro 73.8% — — — 83.1%

Opus 4.7 leads or ties on every benchmark Anthropic published — except the two Mythos sits on. Anthropic now ships the second-best model and tells you who's first.

SOURCE · Anthropic Opus 4.7 announcement · Mythos rows from red.anthropic.com Mythos Preview disclosure · Em-dash = no published score for that model on that benchmark
✦   ✦   ✦
Investigation 02
p.09  —  Reporting by Eddie Belaval

Glasswing.

On April 7, twelve organizations agreed to use a model none of their customers can touch. The precedent starts here.

Consortium 12 founders · 40+ in talks
AWS · Apple · Broadcom · Cisco · CrowdStrike · Google · JPMorgan Chase · Linux Foundation · Microsoft · NVIDIA · Palo Alto Networks · Anthropic
↗   Apr 07   ·   Consortium announced

Project Glasswing is the consortium Anthropic announced alongside Mythos Preview's gated release. The founders: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and Anthropic itself. Forty-plus additional organizations are in early conversation. The goal, stated plainly: harden the world's most critical software infrastructure against a class of vulnerabilities that AI can now find faster than humans can patch them.

The financial structure is unusual for a security consortium. $100M in model credits committed by Anthropic to Glasswing participants — meaning the partners pay nothing for the inference. $2.5M donated to Alpha-Omega and OpenSSF. $1.5M to the Apache Software Foundation. The donations matter because they answer the obvious objection: this isn't a private security tool for big vendors. The funding is meant to flow to open-source maintainers who don't have security teams.

↗   Voices

“The window between vulnerability discovery and exploitation has collapsed — what took months now happens in minutes.”

— Igor Tsyganskiy  ·  Microsoft  ·  Source

Mythos has already produced results. The model has identified zero-day vulnerabilities in every major operating system and every major web browser tested against it. The depth is what gets attention from people who've spent careers in this work: the bugs aren't surface-level. Many are subtle, old, and survived prior automated tooling. A 27-year-old OpenBSD bug. A 16-year-old FFmpeg flaw. Linux kernel privilege escalation chains that no published fuzzer has caught. The Firefox exploit Mythos wrote chained four separate vulnerabilities into a sandbox escape — the kind of compound exploit that a senior offensive security researcher might produce in a multi-week engagement.

“The old ways of hardening systems are no longer sufficient.”

— Anthony Grieco  ·  Cisco  ·  Source

The benchmark Mythos is being judged against is CyberGym, an academic vulnerability-reproduction suite. Mythos hits 83.1%. Opus 4.6 — the previous publicly available frontier model — hits 66.6%. The Firefox-specific test is more dramatic: Opus 4.6 produced two working JavaScript shell exploits across several hundred attempts. Mythos produced 181.

Opus 4.6
2
Working exploits across
several hundred attempts
vs
Mythos Preview
181
Same trial.
Same browser.
A 90× ratio in working browser exploits. The capability gap is not "better." It is categorical.
FIG. 02 · SOURCE: PROJECT GLASSWING ANNOUNCEMENT, APR 7, 2026

“This initiative offers a credible path to making AI-augmented security a trusted tool for every maintainer, not just those with expensive teams.”

— Jim Zemlin  ·  The Linux Foundation  ·  Source

The hard question is institutional, not technical. Anthropic has now established an internal precedent: a model that passes evaluations can still be held back from public release on safety grounds. The Glasswing framing is “defensive use only.” But the model exists. The capability exists. The concession in the Opus 4.7 launch chart was that the capability can't be reproduced by you, no matter how much you pay.

The next time Anthropic invokes this precedent, what will the criteria be? What does the second Glasswing look like? When does the cybersecurity rationale extend to other categories — biological, cognitive, financial — and what's the institutional process for making that call? Anthropic has 90 days to publish the first Glasswing progress report. Read it carefully. The methodology will be precedent.

§   §   §
Feature 03
p.12  —  Reporting by Eddie Belaval

The agent stack got serious.

Managed Agents. The advisor tool. The ant CLI. All three shipped inside April 8–9. The era of writing your own loop is closing — Anthropic is taking that work in-house.

Apr 08 Managed Agents public beta · ant CLI launch
Apr 09 Advisor tool public beta · Cowork GA · SDK parity
↗   Primitives

For the past year, every team building agents on Claude has been writing the same code. A sandbox, a tool harness, a context manager, a streaming loop. The interfaces varied. The components didn't. Anyone who's built an agent in production has shipped some version of all four.

April 8 and 9 was the week Anthropic shipped all four primitives. The era of writing your own loop is closing. Anthropic is taking that work in-house.

Managed Agents went into public beta on April 8 (header managed-agents-2026-04-01). Fully managed harness with secure sandboxing, built-in tools, and server-sent event streaming. You create agents and configure containers through the API — what used to be a Kubernetes problem, an IAM problem, and a tool-routing problem becomes one HTTP call. The pricing model is unannounced; the implication is that Anthropic intends to compete with whatever your in-house agent platform looks like, on its own infrastructure.

The advisor tool shipped April 9 (advisor-tool-2026-03-01). The pattern is simple and overdue: pair a faster executor model with a higher-intelligence advisor that provides strategic guidance mid-generation. Long-horizon agentic workloads get close to advisor-solo quality at executor-model cost. If you've been running Sonnet 4.6 for the bulk work and bouncing to Opus 4.6 for the hard turns, the advisor tool just made that pattern a primitive instead of an architecture.

The ant CLI dropped the same day. A command-line client for the Claude API with native Claude Code integration and YAML-based versioning of API resources — skills, agents, deployments. The implication is bigger than the tool: agents are becoming versionable, declarative configurations rather than code. Treat your agent definitions like Terraform.

↗   Composition
python  ·  messages.create advisor-tool-2026-03-01 
# Old: write your own executor + advisor loop
while not done:
    plan = opus.messages.create(model="claude-opus-4-7", ...)
    for step in plan:
        result = sonnet.messages.create(
            model="claude-sonnet-4-6",
            messages=[..., step],
        )

# New: one call, advisor inline
response = client.messages.create(
    model="claude-sonnet-4-6",
    advisor={"model": "claude-opus-4-7", "trigger": "auto"},
    messages=[...],
    extra_headers={"anthropic-beta": "advisor-tool-2026-03-01"},
)

The week's three releases compose. You can run a Managed Agent that uses the advisor tool internally, with the agent definition versioned as YAML through ant. None of this is theoretical — the docs ship with examples that combine all three.

What this means for the people who built their own agent loops: most of that code is now a glue layer. Some of it remains differentiated — your prompt library, your domain-specific tools, your retrieval pipeline. The harness, the sandbox, the streaming, the context window management — those are commodity now. If your competitive moat was the loop, your moat just shrank.

p.15 Shipped. — 00 — Apr 2026
Timeline

21 days.
56 releases.

21/56
03·27
03·28
03·29
03·30
03·31
04·02
04·03
04·04
04·07
04·08
04·09
04·10
04·13
04·14
04·15
04·16
Flagship moment (model, consortium, GA)
Feature · API · SDK release
Patch · internal
Apr 07
Project Glasswing + Mythos Preview (gated)
12-org consortium. $100M in credits. The shadow goes public — as a thing you cannot have.
Apr 08
Managed Agents + ant CLI
Agent harness as a service. Agent definitions as YAML. The loop becomes infrastructure.
Apr 09
Advisor tool + Cowork GA
Executor/advisor composition becomes a primitive. Cowork leaves beta, picks up RBAC + OpenTelemetry.
Apr 14
Sonnet 4 / Opus 4 deprecation · Vas Narasimhan joins LTBT
Retirement clock starts (June 15). Automated Alignment Researchers paper drops the same day.
Apr 16
Claude Opus 4.7 — GA
Same price. 3× vision. xhigh default. Tokenizer change. And the fourth bar of the launch chart.
Cont.
26 versions of Claude Code, v2.1.85 → v2.1.111
Routines. /ultrareview. /team-onboarding. A UI redesign with a sidebar. The CLI becomes the work environment.
Survey 04
p.17  —  Reporting by Eddie Belaval

Claude Code grew a body.

Twenty-six versions in twenty-one days. The CLI is becoming an IDE, and the IDE is becoming a control plane. When does saying “Claude Code is a CLI” stop being true?

Versions v2.1.85 → v2.1.111  ·  Cadence 1.24 releases/day
Versions 26 Point releases, routine adds, subsystem work. Continuous, unspectacular shipping.
Peak day 7 Apr 09, Apr 14, and Apr 16 each saw seven-plus release events.
Free /ultrareview 3 Cloud code reviews per Pro/Max user per month. Parallel analysis.
Routines ∞ Saved configs running on Anthropic's cloud. Agent-that-ships on a schedule.
↗   Read the changelog

Twenty-six versions shipped between v2.1.85 (March 27) and v2.1.111 (today). Most were point releases: bug fixes, env var additions, plugin tweaks. But woven through them is a clear arc — the CLI is becoming an IDE, and the IDE is becoming a control plane.

Routines are the most consequential addition (v2.1.101 area). A routine is a saved Claude Code configuration — prompt, repos, connectors — packaged once and run automatically on Anthropic's cloud. If you've been wondering when “agent that ships code on a schedule” stops being a custom build, the answer is: now. The infrastructure is hosted. The scheduling is configurable. The cost model is “you pay for the inference” — same as your interactive sessions.

The UI redesign (v2.1.110-111) is the most visible shift. Multiple Claude sessions side by side in one window, with a sidebar. Integrated terminal. File editing. HTML and PDF preview. A faster diff viewer. Anthropic is not subtle about the trajectory. The CLI is becoming the work environment.

/ultrareview (v2.1.111) is comprehensive parallel-analysis code review using cloud compute. Pro and Max users get three free per month. The implication: code review is now a heavyweight LLM task that shouldn't run on your local terminal — it should run on Anthropic's compute and stream back. A pattern other heavy commands will follow.

/team-onboarding (v2.1.101) generates a teammate ramp-up guide from your local Claude Code usage. It's a small feature with a quietly large implication: your AI usage patterns are themselves documentation. Future hires read your transcript shape to learn how to use the tools.

The closing question, after twenty-six versions in twenty-one days: when does Claude Code stop being a CLI? The answer is not far. The diff viewer is faster than your IDE's. The terminal is integrated. The file editor is in the sidebar. The cloud routines are running while you sleep. At some point — probably this year — saying “Claude Code is a CLI” will be technically true and effectively obsolete.

✦   ✦   ✦
p.18 Shipped. — 00 — Apr 2026
Also shipped
↗   Cowork GA

Cowork shipped, finally.

Claude Cowork went generally available on macOS and Windows on April 9, ending what felt like a long beta. The headline is GA itself. Cowork is Anthropic's bet on AI-native team collaboration — not a shared document with AI features bolted on, but a shared agent that works alongside the team and persists across members.

Three additions made the GA more than a milestone. Role-based access controls for Enterprise plans mean admins can carve up which teams get which Cowork capabilities. The pattern is familiar from any enterprise SaaS, but the implication for AI access control is new — what does it mean to scope an autonomous agent to one department's data and not another's? Cowork is the first product that has to answer.

Cowork analytics in the API exposes engagement and adoption data programmatically. OpenTelemetry support is the second half of the same move — Cowork activity becomes telemetry like any other production system. Whether teams use Cowork as a persistent collaborator, not a chat sidebar, depends on whether teams reorganize their workflows to give the agent something to persist. Two quarters from now, we'll know whether Cowork is Slack's AI replacement or Microsoft Bob.

Two papers worth your Friday.

Automated Alignment Researchers (April 14). Anthropic on using LLMs to scale scalable oversight. The paper benchmarks the technique. Read it next to the Mythos disclosure: the same week the lab publicly held a model back for safety, they're publishing on how to use models to do the safety work faster. The two are not unrelated. If you're scaling capability, you have to scale oversight at the same rate or the gap becomes the story.

Emotion Concepts in Large Language Models (March, on transformer-circuits.pub). Researchers found internal representations of emotion concepts in Claude that generalize across contexts and behaviors. The wrong question is whether the models have emotions. The right question is whether they have coherent internal representations of emotions that change what they output. The answer the paper supports is yes.

Both papers point at the same underlying program: the lab is trying to do interpretability and oversight at the speed of capability. Whether they can keep pace is the question that matters most this decade.

Term of the Issue  ·  p.19  ·  First observable 2026-04-16
shadow
release/SHA-doh  ree-LEES/  ·  noun

A pattern in which a frontier AI lab announces a flagship model while simultaneously revealing — through a benchmark chart, a press mention, or an explicit acknowledgment — that a more capable internal model exists but will not be made available to the public.

The shipped model sets the floor of capability buyers can access. The shadow model reveals the ceiling of capability that exists. The space between is where institutional decisions about AI deployment now live.

First observable instance: Anthropic's Opus 4.7 launch (April 16, 2026), in which the announcement chart placed Mythos Preview as the top bar — taller than every released model from Anthropic or its competitors — and noted that Mythos remains gated to security partners under Project Glasswing.

Usage “We're shipping the executor model and shadow-releasing the planner.”
p.20  ·  Notes

Quiet on
the wire.

Mythos broad release — no date, no signal, no roadmap visible. Project Glasswing's first 90-day report will be the next read on whether the gating posture holds. The Bedrock Messages API research preview is us-east-1 only; regional rollout unannounced.

Sonnet 4 and Opus 4 retire from the API on June 15, 2026 — if you haven't migrated, the clock started two days ago. Sonnet 4.5's 1M-context beta dies April 30. The Long-Term Benefit Trust appointed Vas Narasimhan to the Board of Directors on April 14; the pharma-and-policy axis of governance is worth watching over the next quarter for biotech and regulated-industry deal flow.

Notable absences: no fast-mode preview for Opus 4.7 yet, no Haiku 5 signal, no Opus 4.7 on Vertex AI's edge regions, no public roadmap for the ant CLI's Windows binary. All gaps. All worth watching.

The Close

Three weeks. Fifty-six releases. One model you can buy, one you can't.

The shadow is the news.

Build accordingly.

✦
Don't miss the next issue

Get Shipped. in your inbox every Friday at 9 AM ET.

One read on what Anthropic shipped this week. Editorial up front, comprehensive Release Log in back. No spam, unsubscribe one click.

Next issue — Apr 24, 2026
Back of Book  ·  p.21 →

The Release Log

A 1:1 mirror of every Anthropic release in the window. Use it as reference. Share it with your team. Seven sections — A through G — fifty-six entries.

A
Models
2 Models in window

The flagship shipped — and the launch chart conceded a stronger model in the vault.

2026 · 04 · 16 Model

Claude Opus 4.7 — generally available

Anthropic's most capable generally available model. Coding, vision, and self-verification gains over 4.6. New xhigh effort level slots between high and max. Vision processes images up to 2,576px on the long edge (3× prior). Tokenizer change — same input may consume 1.0–1.35× more tokens.

How to use client.messages.create(model="claude-opus-4-7", effort={"level": "xhigh"}, messages=[...])  ·  Pricing unchanged: $5 / $25 per MTok. Available on Claude API, Bedrock, Vertex, Foundry, GitHub Copilot. Re-tune prompts; stricter instruction-following may surprise legacy code.
Why it mattersFirst flagship where Anthropic explicitly conceded a stronger internal model (Mythos) exists.
2026 · 04 · 07 Model · Gated

Claude Mythos Preview (gated)

Unreleased frontier model with strikingly capable cybersecurity performance. 83.1% on CyberGym (vs. Opus 4.6's 66.6%). Found zero-days in every major OS and browser including a 27-year-old OpenBSD bug.

How to useInvitation-only, defensive cybersecurity work only, via Project Glasswing. Not available on the public API.
B
API & Platform
7 Release notes

Managed Agents and the advisor tool entered public beta. Sonnet 4 + Opus 4 retire June 15, the 1M-context beta dies April 30.

2026 · 04 · 16 API

Opus 4.7 API release + tokenizer change

Includes API breaking changes vs. Opus 4.6. New tokenizer means token counts shift 1.0–1.35×. Migration guide published.

How to useRead the Opus 4.7 migration guide before upgrading. Audit your prompt tokenizer counts before swapping model IDs.
2026 · 04 · 14 Deprecation

Sonnet 4 / Opus 4 deprecation

claude-sonnet-4-20250514 and claude-opus-4-20250514 retire June 15, 2026.

How to useMigrate Sonnet 4 → claude-sonnet-4-6, Opus 4 → claude-opus-4-7. After June 15 these model IDs return errors.
2026 · 04 · 09 API · Beta

Advisor tool — public beta

Pair a faster executor with a higher-intelligence advisor that provides strategic guidance mid-generation. Long-horizon agent runs approach advisor-solo quality at executor cost.

How to useAdd beta header advisor-tool-2026-03-01 to requests.
2026 · 04 · 08 API · Beta

Claude Managed Agents — public beta

Fully managed agent harness — secure sandboxing, built-in tools, SSE streaming, container configuration via API.

How to useAdd beta header managed-agents-2026-04-01. Create agents and run sessions via the new endpoints.
2026 · 04 · 08 API · CLI

ant CLI — launch

Command-line client for the Claude API with native Claude Code integration and YAML-based versioning of API resources (skills, agents, deployments).

How to useInstall per the CLI reference. Script API resource lifecycle outside the console.
2026 · 04 · 07 API · Preview

Messages API on Amazon Bedrock (research preview)

First-party Claude API request shape now available on Bedrock at /anthropic/v1/messages endpoint. Runs on AWS-managed infrastructure with zero operator access.

How to useAvailable in us-east-1 only. Contact your Anthropic account executive for access.
2026 · 03 · 30 API

Message Batches API max_tokens raised to 300k

For Opus 4.6 and Sonnet 4.6 only. Long-form content, structured data, large code generation now supported in single batch turns.

How to useAdd beta header output-300k-2026-03-24 to batch requests.
C
Claude Code
26 versions  ·  v2.1.85 → v2.1.111

Twenty-six versions in twenty-one days. Routines, the UI redesign, /ultrareview cloud reviews, /team-onboarding, push notifications, PowerShell on Windows.

2026 · 04 · 16 Code

v2.1.111

Opus 4.7 xhigh available, Auto mode for Max subscribers, interactive /effort slider, “Auto (match terminal)” theme, /less-permission-prompts skill, /ultrareview cloud code review, PowerShell rollout on Windows, Plan files named after user prompts, Ctrl+U / Ctrl+Y clear/restore input buffer.

How to useclaude update. Run /ultrareview in any session for parallel-analysis cloud code review. /effort opens the slider for reasoning depth.
2026 · 04 · 15 Code

v2.1.110

/tui flicker-free fullscreen rendering, push notification tool for mobile alerts, Ctrl+O toggles transcript verbosity, /focus toggles focus view, /plugin Installed tab with priority sorting, session recap for telemetry-disabled users.

How to useRun /tui for fullscreen mode. Configure push notifications via /config. Use /focus to hide non-essential UI during deep work.
2026 · 04 · 14 Code · Cluster

v2.1.109 · v2.1.108 · v2.1.107

Rotating progress hint on extended thinking. ENABLE_PROMPT_CACHING_1H env var for 1-hour cache TTL. Recap feature for session context. Model can discover and invoke built-in slash commands via Skill tool. /undo as alias for /rewind.

How to useexport ENABLE_PROMPT_CACHING_1H=1 to enable 1-hour cache. Use /undo interchangeably with /rewind.
2026 · 04 · 13 Code

v2.1.105

path parameter added to EnterWorktree tool, PreCompact hook with blocking capability, plugin background monitor support via manifest key, /proactive alias for /loop, 5-minute abort timeout for stalled API streams.

How to useConfigure PreCompact hook in .claude/settings.json to block compaction. Plugins can declare monitors in their manifest for background processes.
2026 · 04 · 09 Code

v2.1.101

/team-onboarding generates teammate ramp-up guide from your local usage, OS CA certificate store trusted by default, /ultraplan auto-creates cloud environment, improved brief mode retry logic.

How to useRun /team-onboarding to produce a markdown ramp-up doc tailored to your codebase usage patterns.
2026 · 04 · 09 Code

v2.1.98

Interactive Vertex AI setup wizard, CLAUDE_CODE_PERFORCE_MODE env var for Perforce support, Monitor tool for streaming background script events, subprocess sandboxing with PID namespace on Linux.

How to useRun /setup-vertex for the wizard. Set CLAUDE_CODE_PERFORCE_MODE=1 if you use Perforce. Linux users get tighter subprocess isolation by default.
2026 · 04 · 08 Code

v2.1.97

Focus view toggle in NO_FLICKER mode, refreshInterval status line setting for periodic command re-runs.

How to useSet refreshInterval in your status line config to auto-refresh expensive checks.
2026 · 04 · 02 → 04 · 04 Code · Cluster

v2.1.92 → v2.1.96

forceRemoteSettingsRefresh fail-closed remote config, interactive Bedrock setup wizard, per-model and cache-hit breakdown in /cost, /release-notes as interactive version picker. Mantle-routed Bedrock via CLAUDE_CODE_USE_MANTLE=1. Default effort bumped to high for most tiers. Fixed Bedrock auth failures with AWS_BEARER_TOKEN_BEDROCK.

2026 · 03 · 27 → 03 · 31 Code · Cluster

v2.1.85 → v2.1.91

Opening week. CLAUDE_CODE_MCP_SERVER_NAME/URL, conditional if field for hooks, session-id headers, .jj and .sl excluded from VCS scans. Cowork Dispatch message delivery fixed. MCP tool result persistence overrides. disableSkillShellExecution locks down inline shell. Plugins can ship binaries under bin/. /powerup feature lessons. Rate-limit infinite loop fixed. “defer” permission decisions for PreToolUse hooks in headless mode. CLAUDE_CODE_NO_FLICKER=1 for alt-screen rendering. PermissionDenied hook after auto-mode denials. Named subagents in @ mentions.

D
Claude Apps
6 App releases

Cowork went generally available on macOS and Windows after a long beta. Computer use opened in research preview to Pro and Max plans.

2026 · 04 · 16 Apps

Opus 4.7 in Claude apps

Opus 4.7 available across web, mobile, and desktop with vision and SWE improvements.

How to useUpdate mobile/desktop. Web auto-uses latest. Select Opus 4.7 in the model picker.
2026 · 04 · 09 Apps

Claude Cowork — GA on macOS and Windows

Cowork out of beta. RBAC for Enterprise plans, OpenTelemetry support, Cowork analytics in the Analytics API, custom roles per group.

How to useEnterprise admins: configure groups and assign roles in Console. Pipe Cowork telemetry to your existing observability stack via OTel.
2026 · 03 · 23 Apps · Preview

Computer use research preview for Pro / Max

Claude can access screens and perform tasks independently. Plus Claude Code Dispatch improvements.

How to usePro/Max users opt in to the computer use preview from settings. Read the computer use guide before granting screen access.
2026 · 03 · 12 · 03 · 11 Apps · Cluster

Inline charts · Claude for Excel and PowerPoint enhancements

Claude creates custom charts and diagrams inline in responses on web. Excel and PowerPoint add-ins share conversation context across docs, support skills, and connect to Bedrock / Vertex AI / Foundry via LLM gateway.

How to useUpdate the add-ins. Skills authored in Console show up inside Excel/PowerPoint.
E
Agent SDKs  ·  Python + TypeScript
20 SDK releases

Twenty SDK releases — ten Python, ten TypeScript. Opus 4.7 support, distributed tracing helpers, and a security update worth installing.

2026 · 04 · 16 SDK · PY

claude-agent-sdk v0.1.60

list_subagents() and get_subagent_messages() session helpers, W3C trace context propagation to CLI subprocess (OpenTelemetry), cascading session deletion. Bug fix: setting_sources=[] no longer silently dropped. Bundled CLI to v2.1.111.

How to usepip install -U "claude-agent-sdk[otel]" for distributed tracing.
2026 · 04 · 16 SDK · TS

@anthropic-ai/claude-agent-sdk v0.2.111

Opus 4.7 support, mcp_set_servers per-tool permission_policy for HTTP/SSE servers, startup() and WarmQuery now public API, options.env overlays process.env instead of replacing.

How to usenpm install -D @anthropic-ai/claude-agent-sdk@latest. Use WarmQuery for pre-warmed sessions.
2026 · 04 · 10 SDK · TS · Security

@anthropic-ai/claude-agent-sdk v0.2.101

Security: bumped @anthropic-ai/sdk to ^0.81.0 and @modelcontextprotocol/sdk to ^1.29.0 (GHSA-5474-4w2j-mq4c). Fixed Windows resume-session temp directory leak. Fixed MaxListenersExceededWarning with 11+ concurrent query() calls.

How to useSecurity update — upgrade. npm update @anthropic-ai/claude-agent-sdk.
2026 · 04 · 09 SDK · PY · Cluster

claude-agent-sdk v0.1.57 · v0.1.58

exclude_dynamic_sections option on SystemPromptPreset for cross-user cache hits, "auto" PermissionMode, fixed thinking={"type":"adaptive"} mapping. Bundled CLI to v2.1.97.

How to useUse exclude_dynamic_sections=True to maximize cache hits across users with different dynamic content.
2026 · 03 · 29 → 04 · 04 SDK · PY · Cluster

claude-agent-sdk v0.1.52 → v0.1.56

get_context_usage() on ClaudeSDKClient. typing.Annotated parameter descriptions in JSON schema. ToolPermissionContext exposes tool_use_id and agent_id. session_id option on ClaudeAgentOptions. Fixed --setting-sources empty string when not provided. Fixed query() deadlock with string prompt + hooks/MCP. Fixed silent truncation of MCP tool results >50K chars.

2026 · 03 · 27 SDK · PY

claude-agent-sdk v0.1.51

fork_session(), delete_session(), offset-based pagination, task_budget option, SystemPromptFile support, AgentDefinition gets disallowedTools / maxTurns / initialPrompt. Plus 10+ fixes to async generator cleanup, MCP tool handling, env filtering, process cleanup.

How to useUse fork_session(session_id) to branch a session. Pass task_budget={"input_tokens": N} to cap a run.
2026 · 04 · 08 → 04 · 15 SDK · TS · Cluster

@anthropic-ai/claude-agent-sdk v0.2.97 → v0.2.110

Parity with Claude Code v2.1.97 through v2.1.110. SDKStatus includes 'requesting'. system/memory_recall event. memory_paths on system/init. unstable_v2_createSession now respects cwd / settingSources / allowDangerouslySkipPermissions. New shouldQuery: false field on SDKUserMessage to append without triggering a turn.

How to useSend { shouldQuery: false } to inject context without burning a model turn. Subscribe to system/memory_recall events to see when the agent reads from memory.
F
Research & Publications
2 Papers

Two papers worth a Friday — automated alignment researchers, and the internal representations of emotion concepts in Claude.

2026 · 04 · 14 Research

Automated Alignment Researchers (paper)

“Using large language models to scale scalable oversight.” Anthropic on having LLMs perform alignment work, with benchmarks for whether the technique holds.

How to useRead the paper if you work on agent guardrails. Citations are useful for justifying LLM-based eval pipelines.
Why it mattersSame week Anthropic held Mythos back for safety, they're publishing on automating the safety work itself. The two are connected.
2026 · 03 Research

Emotion Concepts in Large Language Models (transformer-circuits.pub)

Internal representations of emotions in Claude that generalize across contexts and shape behavior.

Why it mattersThe wrong question is whether the models have emotions. The right one is whether they have coherent internal representations that change output.
G
News & Partnerships
2 Items

Glasswing launched with twelve founding orgs and $100M in model credits. Vas Narasimhan joined the Long-Term Benefit Trust board.

2026 · 04 · 14 News

Vas Narasimhan appointed to Long-Term Benefit Trust Board

The LTBT appointed Vas Narasimhan (former Novartis CEO) to the Board of Directors.

Why it mattersAdds pharma-and-policy weight to Anthropic governance. Worth watching for biotech and regulated-industry deal flow.
2026 · 04 · 07 News

Project Glasswing launched

12-organization consortium using Claude Mythos Preview to harden critical software infrastructure. Founders: AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, JPMorgan Chase, Linux Foundation, Palo Alto Networks, Broadcom, Anthropic. $100M in model credits. $4M to OpenSSF, Apache, Alpha-Omega.

How to useOpen-source maintainers of critical infrastructure can apply for Mythos access via Glasswing partners. Participation by application.
Why it mattersFirst time Anthropic explicitly held a model back, post-evaluation, on safety grounds that survived peer review. Sets a precedent.
Sources & Bibliography 40+ citations · Every claim traceable

Every URL consulted in the reporting of Issue 01. If you find a missing citation, file it back via /shipped --revise.

Anthropic — Primary

Release Notes & Docs

Research & Engineering

Coverage & Analysis

Quote Attributions

  • Igor Tsyganskiy, Microsoft — vulnerability discovery to exploitation collapsed. Source: Glasswing announcement.
  • Anthony Grieco, Cisco — old ways of hardening systems no longer sufficient. Source: Glasswing announcement.
  • Jim Zemlin, Linux Foundation — credible path to AI-augmented security for every maintainer. Source: Glasswing announcement.

Benchmark Sources

  • SWE-bench Verified, Pro — Anthropic Opus 4.7 announcement chart
  • GPQA Diamond — Anthropic + competitor publications
  • MCP-Atlas, CursorBench, XBOW Visual-Acuity, BigLaw Bench — Anthropic Opus 4.7 announcement
  • CyberGym — Mythos disclosure / academic vulnerability-reproduction suite
  • OSS-Fuzz — Google open-source fuzzing project
TOTAL SOURCES CONSULTED: 40+  ·  ALL CLAIMS TRACEABLE  ·  REVISIONS WELCOME