The Week in Cloud & AI: Sonnet 5's Million-Token Context, Copilot Vision GA, and the Model Turnover Problem

This week in Cloud & AI

The model ecosystem turned over this week. Anthropic shipped Claude Sonnet 5 with a million-token context window and a new tokenizer, with Fable 5, Mythos 5 and Opus 4.8 reaching general availability in the same platform wave - while GitHub Copilot deprecated Gemini 2.5 Pro and 3 Flash, and GitHub Models moved closer to its July 30 retirement. Models now launch and retire at infrastructure cadence. The second theme: AI features graduating from preview to production - Copilot Vision, TimesFM forecasting inside BigQuery and Claude's background agents all hit GA. And underneath the headlines sat a quieter set of operational notices - GitHub restricting public API endpoints, AWS service sunsets, and five SDK and API breaking changes across Databricks, Terraform and Anthropic - that deserve a slot in your next planning session.

Top stories

Claude Sonnet 5: a million tokens of context - and a new tokenizer that changes your bill

Anthropic launched Claude Sonnet 5 with a 1M-token context window and a redesigned tokenizer that counts roughly 30% more tokens for the same text. Migration from Sonnet 4.6 isn't drop-in: adaptive thinking behaves differently and manual extended thinking has been removed. Introductory pricing runs until August 31, 2026, without Priority Tier support. The same platform wave brought Fable 5, Mythos 5 and Opus 4.8 to general availability. (Release notes)

Why it matters: the context unlock is the headline, but the tokenizer is the operational story - token budgets, rate limits and cost dashboards all shift ~30% without any code change on your side.

Key takeaways:

• Audit prompt budgets and cost alerts before migrating, not after
• Test adaptive-thinking behaviour against your workloads; manual extended thinking is gone
• The introductory pricing window (to Aug 31) is the natural migration deadline

Who should care: AI Engineers, Software Engineers, Engineering Managers · Impact: High

GitHub Copilot Vision is generally available

Copilot now accepts images and PDFs attached directly to chat prompts - across VS Code, github.com and the Copilot CLI, for all subscribers, with no policy changes required. Screenshots of stack traces, architecture diagrams and design PDFs become first-class context alongside code. (Changelog)

Why it matters: multimodal context is becoming a standard capability of coding assistants rather than a demo - "paste a screenshot instead of describing it" is now production-supported in the tool most teams already run.

Key takeaways:

• Works everywhere Copilot chat does; nothing for admins to enable
• Attach error screenshots and design PDFs instead of transcribing them

Who should care: Software Engineers, Platform Engineers · Impact: Medium

TimesFM forecasting lands in BigQuery ML

BigQuery ML now runs Google's pre-trained TimesFM time-series model natively: AI.FORECAST and AI.DETECT_ANOMALIES are generally available, and reachable from Connected Sheets. Forecasting and anomaly detection run directly on warehouse data - no export step, no model pipeline to stand up. (Release notes)

Why it matters: this collapses a whole class of "small forecasting service" infrastructure into a SQL function. For capacity planning, cost anomaly detection and demand forecasting, the warehouse is now the model host.

Key takeaways:

• Zero-infrastructure forecasting on data already in BigQuery
• Worth evaluating against any bespoke forecasting microservice you maintain

Who should care: Data Engineers, Platform Engineers · Impact: Medium

Breaking changes

• Terraform (AWS provider v6.53.0) - breaking changes to Pinpoint SMS Voice V2 phone-number management, bundled with new Bedrock resources. Affects modules that manage SMS resources; the risk is plan-time surprises after a routine version bump. (Release)
• Databricks (SDKs, three languages in one week) - Go v0.152.0 changes query-parameter serialization (ForceSendFields now sends explicit zero values), Python v0.120.0 removes the name field from the IAM User object, and Java v0.123.0 carries its own API breaks. If CI auto-bumps these 0.x SDKs: pin, read, then bump.
• Anthropic (Claude API) - a new agent-memory-2026-07-22 beta header changes response ordering and parameter behaviour on the List Memories endpoint, replacing managed-agents-2026-04-01 on July 22, 2026. Agent builders listing memory data should restart pagination when adopting it. (Release notes)

Security updates

A routine but real patch set from Google Cloud: Cloud Service Mesh, GKE, Apigee hybrid (v1.15.5) and Container-Optimized OS all shipped security fixes. None is dramatic; all matter if you run the component - treat them as this sprint's patch window, not next quarter's. Claude Code also shipped security hardening in v2.1.196.

AI highlights

Google's June roundup landed with substance: Gemma 4 12B for on-device and local use, enhanced computer-use automation in Gemini 3.5 Flash, and new image and video models via API - local inference for sensitive workloads, and automation for the software that has no API. (Announcement)

Cloud highlights

AWS: Amazon WorkSpaces for AI agents reached general availability - managed cloud desktops where agents operate existing applications "with the same security and governance as human users," integrating with any agent framework that speaks MCP. Worth evaluating if your agents need to drive applications that will never grow an API.

Attention required

Announcement	Deadline	Recommended action
GitHub Models retirement	July 30, 2026	Migrate to Azure AI Foundry or Copilot before the playground, catalog and inference API go dark
Copilot drops Gemini 2.5 Pro / 3 Flash	July 31, 2026	Move workflows to Gemini 3.1 Pro / 3.5 Flash; admins must enable the new models
Claude List Memories header change	July 22, 2026	Adopt agent-memory-2026-07-22 and restart pagination
Sonnet 4.6 → Sonnet 5 migration	intro pricing ends Aug 31, 2026	Re-baseline token budgets (~+30%); validate thinking-mode changes
GitHub restricting public API endpoints (stargazers, watchers, subscriptions)	rolling	Check integrations for empty responses / 403s (details)
AWS service availability changes and sunsets	per service	Review the sunset list against your estate

Industry trends

Model lifecycle is now infrastructure lifecycle. New flagships from two vendors, and same-week deprecations of the previous generation. If your systems pin model versions, plan migration like a database upgrade, not a surprise.

Agents are getting dedicated infrastructure. WorkSpaces for AI agents gives agents managed desktops with human-grade governance; Claude's background agents commit and open draft PRs autonomously. Vendors are building the operational envelope around agents: identity, governance, workspaces.

AI features are graduating, not just launching. Copilot Vision, TimesFM in BigQuery, Azure's PII NextGen detection, Claude in Chrome - this week's pattern was previews reaching GA, quietly becoming production-supported surface area.

ReleaseBytes insights

Two of this week's threads point the same direction: MCP is becoming the integration standard for agent infrastructure. WorkSpaces for agents supports "any agent framework using MCP" - an AWS product choosing a vendor-neutral protocol as its integration surface. With agents also gaining Git autonomy (Claude) and UI autonomy (Gemini computer use), the next 6-12 months look like the buildout of an agent operations layer: identity, governance, workspaces and protocol-standard interfaces. Meanwhile, the Sonnet 5 tokenizer change is a reminder that in AI platforms, cost governance is a moving target even when your code isn't - token accounting deserves the same monitoring discipline as cloud spend.

If you only read one thing...

Read the Claude Sonnet 5 release notes. Not for the million-token context - for the tokenizer. A ~30% shift in token counting changes budgets, rate-limit headroom and cost forecasts for every workload that migrates, and the migration itself changes thinking-mode behaviour. It's the week's clearest case of an exciting launch carrying an unexciting operational obligation - and the introductory pricing deadline (Aug 31) puts a date on it.

By the numbers

• 196 releases analysed (Jun 29 - Jul 5, 2026)
• 5 breaking changes (Terraform, three Databricks SDKs, Claude API)
• 7 deprecations / retirements
• 6 security updates
• 114 of 196 releases came from AWS (59) and GCP (55) alone

Thanks for reading. All 196 of this week's releases are browsable on ReleaseBytes - or set a release alert and let the ones that matter find you. If this was useful, share it with a colleague.