Claude Fable 5: Anthropic's Most Capable Public Model
Anthropic launched Claude Fable 5 on June 9, 2026, making it the most capable model the company has ever released for general availability. Fable 5 is built on the same Mythos-class architecture as the restricted Claude Mythos 5, with additional safeguards in place for public deployment. On SWE-Bench Pro — a benchmark measuring real-world GitHub software engineering tasks — Fable 5 scores 80.3%, compared to 69.2% for Opus 4.8 and 58.6% for GPT-5.5. On FrontierCode, which tests demanding production-level coding scenarios, Fable 5 reaches 29.3% versus 13.4% for Opus 4.8. Stripe reported the model compressed five months of engineering effort into days, completing a 50-million-line Ruby codebase migration in a single day.
Beyond coding, Fable 5 delivers major improvements in long-context reasoning and vision. It processes millions of tokens while maintaining coherence, performs three times better than Opus 4.8 when given persistent memory access, and excels at extracting precise figures from detailed scientific documents and reconstructing web applications from screenshots alone. In drug design evaluations, Anthropic's internal protein design experts reported roughly a tenfold acceleration, with nine of fourteen protein targets yielding strong drug candidates.
Three safety systems are built in. Cybersecurity classifiers block assistance with vulnerability exploitation and offensive operations, with external red teams finding zero successful jailbreaks across 30 public techniques. Biology and chemistry classifiers prevent misuse of dual-use scientific knowledge, routing high-risk requests to Opus 4.8 instead. Distillation prevention blocks attempts to extract model weights or capabilities for unauthorised replication. Anthropic reports that 95% of Fable 5 sessions run entirely on the model's own responses without triggering any safety fall-back.
Pricing is $10 per million input tokens and $50 per million output tokens — roughly double the cost of Opus 4.8. Fable 5 is included at no extra charge on Pro, Max, Team, and seat-based Enterprise plans through June 22, after which usage credits are required as capacity expands. Claude Mythos 5, the unrestricted counterpart, remains limited to Project Glasswing cybersecurity partners and select biology researchers through a forthcoming Trusted Access Program.
Read more — Anthropic
Cohere North Mini Code: Open-Weight MoE Coding Agent
Cohere released North Mini Code on June 9, 2026, its first model designed specifically for developer and agentic coding workflows. The model uses a sparse Mixture-of-Experts architecture with 30 billion total parameters and 3 billion active parameters per token — achieved by activating 8 of 128 experts per forward pass. The architecture uses interleaved sliding-window and global attention, supports a 128K token context window, and is available under the Apache 2.0 licence in both BF16 and FP8 quantised formats on Hugging Face.
On SWE-Bench Verified (pass@10 metric), North Mini Code achieves 80.2% after the supervised fine-tuning phase, with reinforcement learning from verifiable rewards (RLVR) pushing it further. On Terminal-Bench v2, it starts at 55.1% pass@10 and gains a 7.9 percentage point absolute improvement after RLVR training, reflecting the benchmark's focus on terminal-based agentic tasks — executing shell commands, managing files, and interacting with development toolchains. On the Artificial Analysis Coding Index it scores 33.4, outperforming several significantly larger models.
Cohere's approach with North Mini Code prioritises efficiency: the 3B active parameter footprint means the model runs at much lower inference cost than its 30B total parameter count might suggest. For teams running self-hosted coding agents or integrating a coding model into IDE tooling via the Cohere API or OpenCode platform, this positions North Mini Code as a competitive open-weight option. The model is particularly strong on agentic assignments that require multi-step terminal interactions, automated code editing, and generating documentation from existing code.
Read more — Cohere Labs
Gemini Code Assist: Gemini 3.5 Flash GA and Next Edit Predictions
Google released Gemini 3.5 Flash as generally available to Code Assist users in VS Code and IntelliJ, making the model available for agent mode, chat, and code generation within the IDE. Gemini 3.5 Flash is Google's current flagship for sustained, high-effort agentic and coding tasks, and its GA availability replaces the earlier preview access that required opt-in.
Alongside the model GA, Code Assist ships two new editing features. Next Edit Predictions predicts where in the current file the developer is likely to edit next after accepting a suggestion, surfacing that location proactively rather than waiting for the developer to navigate. This is designed for refactor and fix workflows where a single logical change spans multiple call sites. The "finish changes" feature acts as an ambient pair programmer: it monitors in-progress edits and infers how to complete them, accepting pseudocode, comments, and partially written code as inputs rather than requiring a formal prompt. A third addition, the "outline" feature, automatically generates short English summaries of code blocks in the Outline panel, letting developers navigate large files without needing to read implementation detail.
These features reflect a broader shift in IDE assistant design from prompt-driven to ambient: rather than requiring developers to stop and write a query, the assistant infers intent from what is already happening in the editor. For Java and Kotlin development specifically, these capabilities layer on top of IntelliJ's existing structural awareness, and teams already using Code Assist for review or boilerplate generation will find the suggestion targeting meaningfully more accurate.
Read more — Google Developers
