Edition 6 · April 20–26, 2026

The week the default interface stopped being “chat” and started being “run the work”

GPT‑5.5 hits prod. Agents hit org charts.

OpenAI put GPT‑5.5 and GPT‑5.5 Pro in front of paying ChatGPT and Codex users, published API dollar rates and a refreshed system card, and paired the capability jump with stricter cyber classifiers plus expanded Trusted Access for Cyber in Codex. The same sprint shipped workspace agents for shared, long-running workflows, WebSockets for the Responses API, a Privacy Filter research direction, and clinician-facing ChatGPT upgrades. Anthropic answered the capacity question with Amazon: up to 5 gigawatts, over $100 billion committed to AWS over a decade, and a path to Claude Platform on AWS. Google used Cloud Next to declare the agentic era — Gemini Enterprise Agent Platform, Gemini 3.1 Pro in the stack, TPU 8t / 8i, and Wiz-shaped agentic security.

Terminal-Bench 2.0 (vendor-reported)
82.7%
GPT‑5.5 vs 75.1% for GPT‑5.4 on OpenAI’s announcement table — eval conditions in the system card.
API list pricing (standard)
$5 → $30
Per million input / output tokens for gpt-5.5 in the API — Batch and Flex at half rate per OpenAI’s pricing copy.

01 · OpenAI · Frontier model

GPT‑5.5: intelligence, token thrift, and serving physics

April 23 (with an April 24 API note on the same page): OpenAI positions GPT‑5.5 as its “smartest and most intuitive” general model yet — stronger on agentic coding, computer use, knowledge work, and early scientific loops, while claiming matched per-token latency versus GPT‑5.4 in real-world serving and materially fewer tokens on Codex-style tasks. GPT‑5.5 Pro is pitched for harder, higher-accuracy workloads. Treat benchmark tables as vendor-published: OpenAI cites SWE-Bench Pro at 58.6% for GPT‑5.5 and flags Anthropic’s memorization discussion on that eval.

Coding (public / cited)

SWE-Bench Pro: 58.6% (5.5) vs 57.7% (5.4). Terminal-Bench 2.0: 82.7% vs 75.1%. Expert-SWE internal: 73.1% vs 68.5%.

Computer use

OSWorld-Verified: 78.7% vs 75.0% for GPT‑5.4 — measures autonomous operation of real desktop environments.

Knowledge work

GDPval (wins or ties): 84.9% vs 83.0% for GPT‑5.4 — structured professional task suite.

OpenAI — Introducing GPT‑5.5 OpenAI — GPT‑5.5 system card

02 · OpenAI · Economics

The API price sheet is now part of the launch story

OpenAI states gpt-5.5 will ship in Responses and Chat Completions at $5 / 1M input and $30 / 1M output tokens with a 1M context window, half that for Batch/Flex, and 2.5× for Priority. gpt-5.5-pro is listed at $30 / $180 per million in/out for “even higher accuracy.” The announcement argues total cost can fall versus GPT‑5.4 when fewer tokens and retries are needed — a claim that only holds once you measure your traffic shapes.

gpt-5.5 $5 in · $30 out · 1M ctx · Batch/Flex −50% · Priority ×2.5
gpt-5.5-pro $30 in · $180 out — positioned for maximum accuracy workloads
Verify Pricing moves — cross-check openai.com/api/pricing before provisioning spend caps.
OpenAI — API pricing

03 · OpenAI · Safety & dual-use

Cyber capability High: classifiers, TAC, and defender lanes

OpenAI classifies GPT‑5.5’s cyber and bio capabilities as High under its Preparedness Framework — not Critical for cyber, but a step up from GPT‑5.4 in testing. Mitigations include stricter classifiers for higher-risk cyber patterns (with a note that tuning may feel “annoying” initially), continued Trusted Access for Cyber starting in Codex with more permissive models for verified defenders, and government partnership language for critical infrastructure. This sits in the same policy arc as prior weeks’ defender programs — now tied explicitly to the 5.5 launch.

RLHF — reinforcement learning from human feedback; vendor “safeguards” summaries are not substitutes for your own abuse monitoring and logging.

OpenAI — GPT‑5.5 (cyber section) OpenAI — Trusted access for cyber defense OpenAI — Preparedness Framework (PDF)

04 · OpenAI · Biology risk surface

Bio bug bounty and GeneBench framing

Alongside the model, OpenAI advertises a GPT‑5.5 Bio Bug Bounty and points readers to deeper biology evaluations — including GeneBench scores on the main announcement (e.g. 25.0% vs 19.0% for GPT‑5.4 on the table shown April 23). Dual-use biology is treated as a first-class release axis next to cyber, not a footnote.

OpenAI — GPT‑5.5 Bio Bug Bounty OpenAI — System card

05 · OpenAI · Enterprise agents

Workspace agents: shared Codex-powered employees

April 22: Workspace agents generalize GPTs into organization-scoped automations — powered by Codex in the cloud, schedulable, Slack-capable, and governed by admin controls including a Compliance API for configuration and run visibility. OpenAI positions them for cross-tool workflows (sales qualification, IT policy routing, month-end close assist). Free until May 6, 2026, then credit-based pricing begins — a clear incentive to experiment before the meter spins.

01

Define — Agents sidebar; natural-language workflow capture with tool and skill wiring.

02

Delegate — Long-running cloud execution; optional schedules and Slack deployment.

03

Control — Per-step approvals for sensitive actions; analytics on runs and adoption.

OpenAI — Workspace agents OpenAI Help — Compliance API

06 · OpenAI · Engineering

WebSockets on the Responses API: lower-latency agent loops

April 22 engineering post: OpenAI adds WebSockets support for the Responses API to reduce round-trip overhead for streaming, tool-heavy agent workflows — the kind of traffic pattern workspace agents and Codex-style automations amplify. Pair this with the announcement that Codex helped optimize GPU partitioning heuristics for heterogeneous request batches (OpenAI cites a >20% token generation speedup from that work on GPT‑5.5’s serving story).

OpenAI — WebSockets / Responses API OpenAI — GPT‑5.5 (inference efficiency section)

07 · OpenAI · Research

Privacy Filter: on-device direction for sensitive prompts

April 22 research announcement: OpenAI introduces OpenAI Privacy Filter, described as an on-device open-weight model that runs locally (initially on Apple silicon Macs with NVIDIA RTX GPUs “coming soon”) to detect sensitive personal information before user text reaches cloud models — with a developer preview and SDK roadmap. If it ships broadly, it changes the data boundary story for regulated teams using ChatGPT and API clients.

OpenAI — Privacy Filter

08 · OpenAI · Health

ChatGPT for clinicians: triage of paperwork, not of patients

April 22 product post: OpenAI details clinician-facing improvements in ChatGPT — documentation assist, structured formats, and workflow hooks aimed at reducing administrative load while repeatedly emphasizing professional judgment and appropriate clinical guardrails. The through-line for hospital IT is integration burden and auditability, not model vanity metrics.

OpenAI — ChatGPT for clinicians

09 · OpenAI · Multimodal

ChatGPT Images 2.0

April 21: OpenAI rolls ChatGPT Images 2.0 — positioned as higher fidelity and more controllable image generation inside ChatGPT, adjacent to the same week’s enterprise and agent pushes. For creative orgs, the policy question remains usage rights and provenance labeling, not peak saturation scores.

OpenAI — ChatGPT Images 2.0

10 · OpenAI · Go-to-market

Scaling Codex to enterprises worldwide

April 21 company post: OpenAI frames Codex as an enterprise surface — SSO, SCIM, audit logs, regional data handling, and admin controls — aligned with the same week’s GPT‑5.5 and workspace-agent narrative. The subtext is procurement: security questionnaires now follow coding agents into the same slots as IDEs and CI systems.

OpenAI — Codex enterprise scaling

11 · Anthropic · Compute contract

Five gigawatts, Trainium cadence, and Claude on the AWS bill

April 20: Anthropic expands with Amazon for up to 5 GW of capacity for training and serving Claude, including Trainium2 ramping first half of 2026 and nearly 1 GW combined Trainium2/3 by year-end. Anthropic states more than $100 billion committed to AWS technologies over ten years, over one million Trainium2 chips in use today, and a forthcoming Claude Platform on AWS (same account and billing). Amazon invests $5 billion now with up to $20 billion additional; Anthropic cites >$30B run-rate revenue and consumer-tier reliability pressure as context.

5 GW
~1 GW

Illustrative bar lengths for “ceiling vs 2026 tranche” — exact phasing is in Anthropic’s primary post, not this graphic.

Anthropic — Amazon compute expansion AWS — Anthropic on Bedrock

12 · Anthropic · Japan footprint

NEC: first Japan-based global partner at 30,000 seats

April 24: NEC will deploy Claude to roughly 30,000 NEC Group employees worldwide as Anthropic’s first Japan-based global partner, with joint industry products for finance, manufacturing, local government, and cybersecurity — including SOC integrations and NEC BluStellar Scenario roadmaps featuring Claude, Claude Code, and Cowork.

Defense in depth

SOC services and “next-generation” NEC cybersecurity offerings get Claude-assisted workflows — read as enterprise upsell, not consumer chat.

Engineering factory

Center of Excellence plus Anthropic-led enablement; internal adoption precedes external SKUs (“Client Zero”).

Regulatory tone

NEC leadership stresses safety, reliability, and public-sector quality bars — language that matches Japan’s procurement culture.

Anthropic — NEC collaboration NEC — BluStellar Scenario

13 · Anthropic · Civic integrity

Election safeguards update

April 24: Anthropic publishes an election safeguards update — framing how Claude should behave around U.S. midterms and other major elections in 2026. Technical readers should pair this with product surfaces (Claude.ai, API policies, and third-party app chains) rather than treating it as purely comms.

Anthropic — Election safeguards

14 · Google Cloud · Conference stack

Cloud Next ’26: agent platform, Gemini 3.1, TPUs, data cloud

April 24 recap: Google positions Gemini Enterprise Agent Platform as end-to-end build/govern/scale for agents, with Gemini 3.1 Pro, Flash Image (“Nano Banana 2”), Lyria 3, and expanded model choice including Claude Opus 4.7. The Gemini Enterprise app adds no-code Agent Designer, long-running cloud agents, and an Agent Inbox. Infrastructure highlights include TPU 8t (training) and TPU 8i (inference) with an 80% better performance-per-dollar claim for 8i, plus Virgo networking fabric and Managed Lustre throughput figures.

Agentic Data Cloud

Knowledge Catalog for autonomous tagging; Cross-Cloud Lakehouse on Apache Iceberg to query data in place — including AWS — without wholesale migration.

Workspace Intelligence

Ask Gemini in Chat across Docs, Drive, Meet, Gmail — action from chat without context hopping.

Model menu

Opus 4.7 on Vertex signals multi-vendor agent runtimes are now marketing table stakes, not science projects.

Google — Next ’26 recap Google Cloud — Gemini Enterprise Agent Platform Google Cloud — TPU 8t / 8i deep dive Google Workspace — Workspace Intelligence

15 · Google Cloud · Security

Wiz + Google: agentic defense lanes

Google’s Next security narrative pairs threat intelligence with Wiz (now inside Google Cloud) — specialized agents for threat hunting, detection engineering, and third-party context. The honest operational read: agents that write rules autonomously still need human ownership of blast radius and change management.

Threat Hunting agent

Proactive hunts and autonomous rule drafting — verify outputs against your detection-as-code standards.

Detection Engineering agent

Coverage-gap analysis and new detection creation — watch for duplicate logic and alert fatigue.

Google Cloud — Security for the AI era Google Cloud — Wiz acquisition

16 · Policy · Compliance rail

EU AI Act clock beside your agent registry

As Google sells an Agent Registry and OpenAI sells workspace agents with compliance APIs, EU operators still map obligations to the Commission’s official implementation timeline — provider vs deployer vs GPAI systemic risk categories phase through 2026–2027. The design pattern this week is convergent: governed automation is the SKU.

GPAI — general-purpose AI; DPIA — data protection impact assessment; agent memory + cross-SaaS tools make both salient.

EU AI Act Service Desk — Timeline

17 · Steel man

What if the agent wave mostly centralizes spend?

Optimistic read: Workspace agents and governed cloud agent platforms compress integration time — Rippling-style wins without standing up a bespoke agent stack.

Skeptical read: List prices for GPT‑5.5 Pro and Priority multipliers concentrate frontier inference spend with a handful of vendors; “token efficient” claims need per-workload proof.

Infrastructure read: Anthropic’s multi-gigawatt AWS deal is capacity insurance — until electrons flow, latency and quota stories stay political.

18 · Forward calendar

The week ahead

Anchors from primary pages cited here — confirm times in local calendars.

May 6, 2026

Workspace agent pricing

Free preview ends; credit-based pricing begins per OpenAI’s workspace agents post.

May 8, 2026

OpenAI macOS certificate rotation

Still on the calendar from the prior edition’s Axios supply-chain response — verify against OpenAI’s latest guidance if you ship desktop clients.

H1 2026

Trainium2 capacity (Anthropic × AWS)

Anthropic cites meaningful Trainium2 compute in the next three months; track Bedrock quotas and regional expansion announcements.

Rolling

GPT‑5.5 API availability

OpenAI’s launch note promises Responses/Chat Completions “very soon” after consumer rollout — watch API changelog and safety gates for your tier.

Rolling

Claude Platform on AWS

Described as “coming soon” with account-team access — enterprise admins should expect pilot queues.

June 8, 2026

Apple WWDC26

Developer conference — likely Apple Intelligence and on-device model roadmap updates adjacent to Privacy Filter’s Mac positioning.

OpenAI — May 6 pricing note OpenAI — May 8 cert rotation Apple — WWDC26
What is Frontier AI Weekly?