OpenCode adds Ring 2.6 1T with 256K context and free limited-time access

• OpenCode added Ring 2.6 1T with reasoning enabled, a 256K context window, and free limited-time access, according to opencode's availability post.
• Kilo also shipped Ring 2.6 1T into its model picker and described it as a trillion-parameter thinking model for long-horizon tasks, coding, and agent workflows in kilocode's launch post.
• Benchmark claims around the model are aggressive: kilocode's chart post says Ring 2.6 1T leads on AIME 26, ClawEval, Gaia2-search, and Tau2-Bench Telecom, while holding close on GPQA Diamond and ARC-AGI-V2.
• The OpenCode rollout landed alongside a fast patch run, and LLMpsycho's v1.14.46 post says the latest release fixed old-session loading, tolerated broken MCP schemas, and patched a Plan Mode bypass.

You can open the OpenRouter listing, check the Kilo CLI page, and skim the OpenCode v1.14.46 release. The weirdly useful detail is that this small model-picker update came with a free-access window on both OpenCode and Kilo, plus a benchmark graphic that places Ring 2.6 1T against GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.7.

OpenCode's announcement was short and concrete: Ring 2.6 1T is live, text-only, supports reasoning, and gets a 256K context window. That makes the story less about a new interface surface and more about a new high-end model option inside the existing editor.

A follow-up post from dingyi's OpenCode mention echoed the same limited-time free access. The only linked surface in the evidence is the OpenRouter listing, so the clearest read is that OpenCode is exposing the model through its provider stack rather than introducing a Ring-specific workflow.

Kilo pitched the same model more explicitly as a flagship thinking model for production work. The launch image in kilocode's launch post names four target workloads:

• Long-horizon tasks
• Coding and engineering
• Complex reasoning
• Agent workflows

The benchmark card in kilocode's chart post claims Ring 2.6 1T leads on AIME 26, ClawEval, Gaia2-search, and Tau2-Bench Telecom. The same chart shows weaker placement on SWE-Bench Verified, where Ring scores 74.0 against 87.6 for Claude Opus 4.7, and on GPQA Diamond, where it trails Gemini 3.1 Pro and Claude Opus 4.7.

Kilo also said in kilocode's weekly thread that Ring 2.6 1T shot to the top of its internal leaderboard right after launch. The post does not explain that leaderboard's methodology, so the more concrete evidence is still the chart plus the day-one availability in Kilo's model picker.

The OpenCode side of this rollout was noisy in a useful way. Three releases shipped in quick succession: v1.14.44, v1.14.45, and v1.14.46.

Across LLMpsycho's v1.14.44 post, LLMpsycho's v1.14.45 post, and LLMpsycho's v1.14.46 post, the fixes cluster around the boring parts that decide whether a new model is actually usable:

• workspace upgrade failures tied to a time_used field
• active-model acceptance and workspace query parameter fixes
• numeric and boolean API parameter fixes
• old session loading restored
• broken MCP schemas tolerated
• a Plan Mode bypass fix

That patch train is the last interesting detail here. OpenCode did not just add Ring 2.6 1T, it was also tightening the editor around session recovery, provider parameter handling, and MCP edge cases at the same moment the new model went live.