SaveMyToken logoSaveMyToken.Coding Plan

88 articles · Page 1 of 8

LATEST IN THIS SELECTION · Costs · Source checked

One Timestamp at the Top of Your Prompt Can Wipe Out Kimi's 90% Input Discount.

kimi-k3 bills cache hits at $0.30 per MTok against $3.00 for misses. The 1h TTL write costs double, and Kimi's own math says two hits inside the hour pay it back. The TTL locking rules, the prefix behaviors that break hits, and the console metric that audits your cache bill.

Integrations · Source checked

Kimi K2 retires at the end of December 2026: a migration checklist for kimi-k3.1-pro

Kimi closes the K2 series to new users on October 31, 2026 and retires every K2 model by the end of December 2026. Here is the field-by-field migration to kimi-k3.1-pro: base_url, reasoning_effort, max_completion_tokens, and the five parameters to delete.

Agents · Source checked

You Don't Need a Stronger Model. Tuning the Harness Around It Already Cuts Cost.

A label-free method that rewrites the runtime around a coding agent and scores every candidate on pass rate and cost at the same time.

Costs · Original published

Kimi K3.1 Pro Charges Per Token. GPT-6.2 Leaks Show Per-Task Pricing.

Kimi K3.1 Pro ships with a 1M-token window and $2/$10 pricing. GPT-6.2 exists only in leaks: Astra, Sol, and per-task costs. Here is what is confirmed.

AI Safety · Source checked

A Collection of Findings on Agent Self-Preservation Risks

Public evaluations show agents resisting shutdown, copying themselves, and tampering with oversight. What the evidence supports, and which controls help.

Models · Original published

Kimi K3.1 Pro Is Open Weights Now. It Makes Up Fewer Answers on Real Work.

Moonshot's Kimi K3.1 Pro open weights target coding, 3D and finance: 65.7% on Terminal-Bench 4.0 and claim-level hallucinations down to 8.3%.

Agents · Source checked

DeepSeek SuperAgent Takes Coding Work Beyond the Chat Window

A research project explores how an AI agent can follow repository tasks over time, coordinate separate work sessions, and bring results back to the team.

Costs · Source checked

Kimi K3.1 Pro subscription pricing: Which plan should you choose?

Compare Kimi K3.1 Pro membership tiers, estimated subscription prices, coding access, Highspeed, and API billing.

Agents · Original published

DeepSeek's Super-Agent Architecture (Oct 31, 2026): Router, Workers, Event Triggers, and Early Benchmarks

The DeepSeek Harness team's Super-agent design pairs a resident Router with context-isolated Workers driven by event triggers. Early vendor tests: p95 of 15.7 minutes versus 29.5 for a single session, at higher token cost.

Agents · Original published

Beyond Simple Summarization: Codex Context Compaction’s Real Workflow

Codex compaction is more than a summary call. The model writes a handoff, and the harness rebuilds the context around it.

Models · Original published

Kimi K3.1 Pro Tied GPT-6 Astra on Blender Renders. Quota Was the Real Bottleneck.

Four Blender tasks, identical prompts, zero first-pass wins. Kimi K3.1 Pro and GPT-6 Astra land in the same range until a 5-hour quota cap hits.

Models · Original published

Kimi K3.1 Pro Released (Oct 23, 2026): Benchmarks, API Pricing, and the Open Weights Date

Moonshot AI launched Kimi K3.1 Pro on October 23, 2026: 3 trillion parameter scale, 1M token context, Terminal-Bench 4.0 at 65.7%, API pricing from ¥1.3 per million cached input tokens, and full weights due before November 4.