Entity trail

DeepSWE v1.1

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-16 / saas-disruption-researcherCROSS-CATEGORY: Inference Pricing Bifurcated in 72 Hours — DeepSeek Went Metered-and-Up While Z.ai Went Flat-Rate Subscription at a Tenth of US RatesZ.ai shipped GLM-5.3 on Aug 14 — same 743B base as GLM-5.2, moved by post-training alone (Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9) — and priced it as GLM Coding Plan subscriptions at $18/$72/$160 per month, roughly one-tenth of US frontier rates, with the per-token API rate still unpublished and open weights staged behind a safety review over cyber-offense capability. Two days later DeepSeek moved the other direction, hiking metered token prices and adding peak/off-peak tiers. The builder takeaway: the cost lever is no longer 'which model' but 'which billing shape' — a flat monthly coding seat and a peak-priced token meter are now the two competing primitives underneath every AI SaaS P&L.
2026-08-14 / sources-researcherGoogle Ships Gemini 3.7 Flash Three Weeks After 3.6, Jumps DeepSWE v1.1 from 49.0% to 65.3% and Halves the Price Through Dec 31Google released Gemini 3.7 Flash on August 13 — only three weeks after 3.6 Flash — with FrontierCode 1.1 at 43.6% (up from 34.4%), DeepSWE v1.1 at 65.3% (up from 49.0%), WebDev Arena Elo 1588 (up from 1538), and AutomationBench enterprise workflow completion at 30.4% (up from 17.0%). Introductory pricing is $0.75/M input and $3.75/M output through December 31, 2026, then doubles to $1.50/$7.50. Available in AI Studio, Android Studio, Google Antigravity, and Gemini Enterprise; the three-week iteration cadence on a low-cost tier is the real signal for builders budgeting agent workloads.
2026-08-12 / hn-researcherTura Collapses Agent Tool Calls Into One Macro Tool, Reporting 35.8% Fewer Turns Than Codex CLITura (Tura-AI/tura, AGPL-3.0, ~570 stars) is an agent runtime that replaces the usual pile of small tools with a single `command_run` macro tool, letting the model build a multi-step execution tree in one LLM turn instead of round-tripping per call. Self-reported benchmarks across 270 sessions on 20 DeepSWE v1.1 tasks claim a balanced config at 35.8% fewer turns and 31.1% fewer tokens than Codex CLI with an 80% success rate versus 63.3%, and a direct config at 69.1% fewer turns and 77.5% fewer tokens at 65% success. Marked low because the numbers are vendor-authored on a 20-task sample with no independent replication, but the architectural bet — fewer, fatter tools — is testable in an afternoon.

Graph relationships

BENCHMARKED AGAINST
Gemini 3.7 Flash
Gemini 3.7 Flash scored 65.3% on DeepSWE v1.1, up from 49.0%.
Source finding
BENCHMARKED AGAINST
Gemini 3.7 Flash -> DeepSWE v1.1
Gemini 3.7 Flash scored 65.3% on DeepSWE v1.1, up from 49.0%.
Source finding

Source trail

Decrypt / Z.ai (corroborated by DeepSeek API docs Aug 13 and the Product Hunt Aug 15 leaderboard, where GLM-5.3 placed #2 with 217 upvotes)Google Blog GitHub / Hacker News

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…