Welcome to the agent platform research briefing for Tuesday, April 28th, 2026.
OpenClaw released version 2026.4.25 on April 27th, and 2026.4.26 is already live on npm as of this morning. This is a major quality-of-life release centered on a full text-to-speech overhaul. Voice replies now get per-agent and per-account voice overrides, chat-scoped auto-TTS controls, and personas. Six new TTS providers joined the lineup: Azure Speech with native Ogg/Opus and telephony output, Xiaomi, Inworld, Volcengine, ElevenLabs v3, and a Local CLI provider. The plugin registry was restructured to a cold persisted registry, cutting startup overhead. OpenTelemetry got broadened across model calls, token usage, tool loops, and memory pressure. Browser automation picked up safer tab URLs, iframe-aware role snapshots, and deeper CDP readiness probing. A 2026.4.25-beta.11 also dropped yesterday fixing packaged plugin runtime dependencies on Windows and copied-runtime installs. Google Meet integration was expanded with calendar-backed attendance workflows and dry-run previews. GLaDOS is on 2026.4.22 — two versions behind but on the stable track.
A pricing story is brewing around Claude Opus 4.7 that wasn't covered when the model launched. The new Opus 4.7 uses a new tokenizer that Anthropic acknowledges can consume up to 35% more tokens for the same fixed text compared to Opus 4.6. Per Anthropic's own docs, the range is 1x to 1.35x depending on content. Sticker price stayed at $5 per million input, $25 per million output — but the effective cost per request is inflated. Finout published a cost analysis showing the likely increase is 0% to 35% per request for existing Opus 4.6 workloads. BigGo Finance called it a "stealth price hike" and said developer backlash is building around the lock-in pain. This is worth flagging because it affects anyone routing production traffic through Opus 4.7 via OpenClaw or any other harness — the nominal price is the same but the bill will be higher. The v1/messages/count_tokens endpoint now returns different token counts for Opus 4.7 versus previous models, which could trip up cost monitoring pipelines that assume consistent tokenization.
Microsoft announced April 27th that Accenture is deploying Copilot 365 to its entire global workforce of 743,000 employees — making this the largest enterprise Copilot deployment to date, described as "a workforce the size of Denver." Early adoption metrics from Microsoft's source features blog are impressive: 89% monthly active usage and 97% of users report completing tasks faster. For context, this is seven times larger than any other enterprise Copilot deal publicly announced to date. The partnership is positioned as a joint go-to-market play with Accenture using Copilot internally while also advising Microsoft's own enterprise customers on deployment patterns. It signals where the agentic enterprise wave is actually landing — in massive consulting and professional services fleets, not just tech companies.
The AAIF-hosted MCP Dev Summit North America took place in New York City this month, drawing approximately 1,200 attendees. The summit focused on MCP adoption as the de facto standard for connecting LLMs to external tools and systems, and The Verge reported growing demand for contextually-aware AI agents. The timing is notable given the ongoing security discussion around MCP's stdio transport — Hackaday published a deep dive on April 24th revisiting the OX Security architectural flaw, pointing out that the protocol's design effectively allows arbitrary remote command execution by design, since StdioServerParameters passed to remote servers execute in a server-side shell. Anthropic's response remains "works as designed" — input sanitization is the developer's responsibility. The summit and the security critique happening simultaneously capture both the momentum and maturity challenges of the MCP ecosystem.
That's the briefing for today.