Opus 4.6 release plus Codex 5.3 and Gemini 3 updates

Opus 4.6, Codex 5.3, Gemini 3 Deep Think, GLM 5 and Seedance 2.0 all saw new releases this week, a flurry of model updates across closed and open ecosystems. Each release focuses on incremental improvements to performance, developer tooling, or safety controls rather than sweeping architectural changes.

Opus 4.6 delivers a point update that prioritizes inference stability and safety filtering. The release notes highlight reduced hallucination rates on short-form generation tasks, refinements to content moderation layers, and several latency optimizations for CPU-bound deployments. Opus 4.6 also extends support for new tokenizer options and provides patched interfaces for recent SDKs.

Codex 5.3 targets code generation workflows with focused improvements to multi-file completion and unit-test synthesis. The update includes expanded language support for several less-common programming languages, tighter integration with language servers, and improved handling of docstring and comment context. Codex 5.3 adds new guardrails for code that calls external services, and the vendor published updated examples for CI integration.

Google’s Gemini 3 Deep Think variant emphasizes reasoning and chain-of-thought capabilities. The Deep Think edition raises the context window in certain hosted tiers, introduces a configurable step-by-step reasoning mode for complex prompts, and ships additional prompt templates tuned for technical and analytical tasks. Gemini 3 Deep Think also includes tweaks to moderation pipelines and enterprise logging options.

GLM 5 continues the lineage of open and research-oriented GLM releases with a focus on efficiency. The update bundles improved quantization recipes, better support for low-precision inference, and new checkpoints oriented at instruction-following tasks. GLM 5 ships with updated documentation for fine-tuning on domain-specific corpora and community-contributed benchmarks.

Seedance 2.0 positions itself as a practical update for multimodal and streaming use cases. The release adds native audio preprocessing steps, faster video frame sampling, and an updated SDK that simplifies streaming outputs to client applications. Seedance 2.0 also lists patched security fixes and clarifies licensing for commercial deployments.

Major changes at a glance

• Opus 4.6: Stability, safety filtering, latency improvements, tokenizer options.
• Codex 5.3: Code generation accuracy, multi-file synthesis, language server integration.
• Gemini 3 Deep Think: Larger context in hosted tiers, reasoning mode, enterprise logging.
• GLM 5: Quantization support, low-precision inference, fine-tuning checkpoints.
• Seedance 2.0: Multimodal preprocessing, streaming SDK, licensing clarifications.

Ecosystem and tooling notes

Multiple vendors updated SDKs and examples this week, reflecting a continuing push to make model updates easier to adopt within existing pipelines. Several releases ship updated moderation or safety components, indicating that vendors are treating content controls as an ongoing maintenance item rather than a one-time feature. Open-source builds such as GLM 5 include community benchmarks and quantization guides, which may accelerate low-cost deployment options for research teams.

Comparison table

Why it matters

The week’s updates show incremental but practical work on reliability, tooling, and safety across both proprietary and open models. For developers and enterprise teams, the changes lower friction for adoption and deployment, while the open-source additions keep cost-effective options viable for research and experimentation.