Fable 5 is back! The frontier labs released new models and model capabilities, e.g., GPT-5.6 series, Sonnet 5, Nano Banana 2 Lite, Gemini Omni Flash, computer use in Gemini 3.5 Flash. Anthropic introduced new enterprise tooling, e.g., an apps gateway, an agent identity approach for multiplayer situations. Cognition launched a new multi-model harness using a new sidekick pattern. OpenAI and Anthropic both released economic studies on how AI is reshaping work. Werner Vogels wrote about culture both adapting and staying true in the midst of technology change.
RELEASES
① Introducing computer use in Gemini 3.5 Flash — June 24, 2026
Summary: Google has integrated computer use as a built-in tool in Gemini 3.5 Flash, as it previously required a dedicated standalone model. Developers can now use 3.5 Flash to build agents that act across browser, mobile, and desktop environments, with two optional enterprise safeguards: 1/ explicit user confirmation for sensitive actions, and 2/ automatic task termination on detected prompt injection.
Signal: Computer use integration into general-purpose models signals that the capability is maturing from experiment to standard primitive. The opt-in nature of safeguards still requires teams to make deliberate choices about when to trust the model to act without a human checkpoint.
② Previewing GPT-5.6 Sol: a next-generation model — June 26, 2026
Summary: OpenAI is launching a limited preview of their GPT-5.6 series, which includes 1/ Sol, their flagship model, 2/ Terra, for everyday work, and 3/ Luna, a fast and affordable model. Sol includes new max and ultra reasoning modes, with ultra using subagents to parallelize complex work. Sol sets the new state of the art on Terminal-Bench 2.1 and is competitive with Mythos Preview on ExploitBench at ~1/3 the output tokens. Pricing runs from $5/$30 to $1/$6 input/output across the three tiers.
Signal: The phased release coordinated with the U.S. government, rather than launching broadly and risking a retroactive suspension like Fable 5 and Mythos 5, suggests that OpenAI internalized the precedent set by that directive. With Fable 5 and Mythos 5 suspended over a narrow jailbreak concern, GPT-5.6 Sol’s arrival as a competitive alternative at ~1/3 the output tokens illustrates how abrupt regulatory intervention creates market openings that move faster than the underlying technical justification.
③ Introducing the Claude apps gateway for Amazon Bedrock and Google Cloud — June 29, 2026
Summary: Anthropic introduced the Claude apps gateway, a self-hosted control plane that connects Claude Code to Amazon Bedrock and Google Cloud without requiring per-developer credentials. It is deployed as a single stateless container, handles identity via any OIDC provider, enforces policy centrally, routes telemetry, enables multi-provider inference routing, and applies spending caps at the organization, group, or user level.
Signal: The gateway addresses major friction points for enterprise Claude Code adoption: 1/ centralized access control, 2/ resiliency, and 3/ cost visibility. The self-hosted gateway, which keeps inference traffic direct to the provider and usage telemetry on customer infrastructure, reflects how seriously data residency and privacy concerns weigh on enterprise decision making.
④ Devin Fusion: Frontier Performance at 35% Lower Cost — June 29, 2026
Summary: Cognition introduced Devin Fusion, a multi-model harness for their Devin coding agent that runs a main frontier agent and a cheaper sidekick agent in parallel. The main agent handles planning, ambiguity, and final review while delegating execution to the sidekick. Meanwhile, lightweight classifiers evaluate work mid-session in case a model switch is required or work needs to shift back to the main agent. On their FrontierCode benchmark, Fusion achieves a 35% cost reduction while matching frontier model performance, rising to 41% with Fable 5.
Signal: The sidekick architecture improves on simple model routing by keeping a frontier model in the loop for judgment calls while delegating execution, avoiding quality degradation that plagues most cost-reduction approaches. The finding that Fable 5 amplifies cost savings more than other models suggests that hybrid harnesses become more effective as base models improve at orchestration.
⑤ Introducing Claude Sonnet 5 — June 30, 2026
Summary: Anthropic launched Claude Sonnet 5, positioning it as the most agentic Sonnet model to date, with performance close to Opus 4.8 at introductory pricing at $2/$10 per million input/output tokens through August 31, rising to $3/$15 thereafter, versus $5/$25 for Opus 4.8. The model improves on Sonnet 4.6 across reasoning, tool use, coding, and knowledge work, with lower rates of hallucination, sycophancy, and misaligned behavior, and better resistance to prompt injection attacks. Notably, Anthropic deliberately did not train Sonnet 5 on cybersecurity tasks and ships it with cyber safeguards enabled by default.
Signal: Sonnet 5 narrows the gap between Sonnet and Opus-class models at a meaningfully lower price point, shifting the cost/performance calculation for agentic deployments. Further, the deliberate decision to withhold cybersecurity training reflects a capability differentiation strategy where Anthropic is actively segmenting what each model tier can do rather than letting capability trickle down uniformly.
⑥ Start building with Nano Banana 2 Lite and Gemini Omni Flash — June 30, 2026
Summary: Google introduced Nano Banana 2 Lite, their fastest and most cost-efficient Gemini image model, and Gemini Omni Flash, for high-quality video generation and conversational editing. Both models are available in Google AI Studio, Gemini API, Gemini Enterprise Agent Platform, and the Gemini application. Combined, they are targeted at building comprehensive, end-to-end multimedia experiences that connect rapid image generation with video creation and editing.
Signal: The joint release signals Google’s push to own the full generative multimedia stack, with tooling to chain the two models together via the Interactions API. The Interaction API’s support for persistent session context across sequential edits represents a shift from generation as a one-shot action to generation as ongoing, stateful workflow.
⑦ Redeploying Fable 5 — June 30, 2026
Summary: Anthropic redeployed Claude Fable 5 on July 1 after U.S. export controls forced a global suspension of both Fable 5 and Mythos 5, triggered by research from Amazon that discovered a narrow jailbreak, which Anthropic noted was reproducible by less capable models. A new classifier now blocks the technique in over 99% of cases, with a tradeoff of more false positives on routine coding tasks. Anthropic outlined four criteria in a scoring framework for jailbreaks: 1/ capability gain, 2/ breadth of capability gain, 3/ ease of weaponization, 4/ discoverability, co-developed with Amazon, Microsoft, Google, and other Glasswing partners.
Signal: No shared standard exists for what constitutes a dangerous AI jailbreak, leaving AI developers and regulators operating on incompatible definitions. The proposed framework and deeper collaboration with the U.S. government signal that self-certified AI safety is ending while a more formalized compliance approach is beginning.
SECURITY
① Agent identity in Claude Tag: a new access model for autonomous, team-wide AI — June 24, 2026
Summary: Anthropic documented their approach to agent identity for Claude Tag, their multiplayer agent for shared team channels. Rather than acting on behalf of individual users, Claude Tag holds its own agent identity, with requisite credentials, repository access, connectors, skills, and standing instructions. The default identity is scoped at the workspace level and overridable per channel. While public channels share a workspace-level identity, private channels get a distinct identity. DMs use the individual user’s own claude.ai account.
Signal: This is a meaningful shift from enterprise access control design, wherein IAM and RBAC frameworks leverage per-user permissions. As agents become multiplayer and longer-running, the per-user delegation model breaks down. Anthropic is moving the identity boundary from person to channel and looking to supplement this change with just-in-time credentials.
RESEARCH
① How agents are transforming work — June 25, 2026
Summary: OpenAI’s economic research paper documents how Codex displaced ChatGPT as the primary AI tool across every department, including legal, finance, and recruiting, in under a year since public release, with non-developer users growing 137x-189x. Nearly a quarter of requests correspond to tasks that would take a person more than an hour, reflecting a shift from single interactions toward delegated, long-horizon work.
Signal: The non-developer adoption numbers are the most consequential finding, indicating that agentic tools are crossing into general knowledge work faster than assumed by workforce planning. Employees are already executing technical tasks outside their job description, which changes the calculus for how organizations redesign business workflows and which skills are most valuable.
② Anthropic Economic Index report: Cadences — June 26, 2026
Summary: Anthropic’s June 2026 Economic Index report tracks how Claude usage is shifting from conversational interactions toward long-running agentic tasks and includes the first results from a linked survey of approximately 9,700 Claude users. Usage mirrors the rhythms of daily life, e.g., work queries dropped on the weekends, news queries are common in the morning, tax queries were 8x as common close to the tax filing deadline. Regarding job expectations, 6 in 10 respondents expect AI to handle a larger share of their work in 12 months, and over a third expect it to handle most or nearly all of it.
Signal: The most counterintuitive finding is that users who delegate the most to Claude via full task automation are the most optimistic about their future pay and employability. This inverts the common framing that heavy AI delegation threatens workers, as the data suggests that those who experience AI’s benefits directly have a more positive outlook on what it means for their career.
ONE TO READ
① A return to two-pizza culture — June 30, 2026
Summary: Werner Vogels revisits Amazon’s two-pizza team culture, wherein small autonomous teams have end-to-end ownership, and discusses how coding agents are prompting a meaningful update to Amazon’s Working Backwards process. Writing was originally an equalizer, making product definition accessible to anyone regardless of technical background. However, now that anyone can prototype a functional product in an evening with a coding agent, Werner proposes inverting the process: 1/ build first, 2/ use it as a customer would, and then 3/ write. The document you produce after building and testing is grounded in reality rather than assumption. Amazon Quick Desktop, which started as an overnight prototype built with Kiro, is the case in point.
Note: I wrote some personal thoughts about this on LinkedIn, digging into the value of writing for clarity of thought.
Always be learning.
heeki reads #7
Written by Heeki Park, Principal SA @ AWS. Opinions are my own.
Alcurio is where alchemy meets curiosity.