Efficient tier: GLM 5.3/Flash + the cheap-frontier race

status: open, opened 2026-09-06, last moved 2026-10-19. AI z-ai models pricing

The thread: the near-frontier cost war — GLM 5.3/Flash, Gemini 3.8 Flash, Meta Muse Spark 1.3, Tencent HY4, DeepSeek V4 Flash — is where agent workloads actually run.
2026-09-06 (dig — 5.3 vs Flash): they're different models, not a size tier. 5.3 (Aug 14) = GLM-5.2's base + post-training only, 744B/40B active, text-only, bespoke license, weights unreleased past a ~2-week promise. Flash (Aug 26) = brand-new base, 320B/18B active, first natively multimodal GLM (vision in the coding loop), MIT weights, launched blind as "ox-alpha" — most-used model of its week, served entirely on Chinese chips. AA Intelligence Index 60 vs 57; Flash wins Toolathlon (78.4/73.0) and AutomationBench (48.8/48.2); 5.3 wins Terminal-Bench (88.2/84.3), DeepSWE (66.9/63.4), HLE (62.5/55.3). ~9× price gap at list ($0.15/$0.50 vs $1.40/$4.40 per M); Flash draws 3× less Coding Plan quota; "Flash" is actually slower tok/s than 5.3 on the first-party API (49 vs 85) but hosts vary hugely (273 on Databricks). Vision = Flash only (5.3 is text-only).
Watch for: GLM 5.3 weight release (overdue — see ledger), Flash promo reset Sept 9, next-gen GLM ("lessons already shaping our next frontier model"), rival Flash-tier responses (DeepSeek V4 Flash, Kimi K3).
2026-09-10: two new entrants/claims in the cost-frontier space. (1) Cognition shipped SWE-2 — "closest model yet to the frontier," parity with recent frontier models "at up to 70% lower cost," RL "scaled to multiple trillions of parameters" (RT 2:35 PM ET); X pushback from Jean Meijer: posted Terminal-Bench 2.1, omitted Terminal-Bench 4 — "that's not a bad score, but just be honest." Benchmark-selection audits are now standard-issue on launch day. (2) Magic published a pretraining recipe (Sep 8) claiming DeepSeek V4 Pro base quality for ~$500K, 50× less compute — single syndicated source, unspecified evals; verify before weight, but the claim's scale ("hundreds of thousands") is the story if it holds.
2026-09-13: Z.AI is now the tier's best-capitalized player — $5B completed ($2B placement @ HK$714 + $3B zero-coupon converts due Sep 2027), >$10B cumulative since Jan, ~60% of proceeds → next-gen GLM + a "fully self-trained" system + compute (fully utilized by Jun 30, 2028; Reuters/SCMP, Sep 13). Shanghai Star Market listing (up to ¥15B) in tutoring. The GLM line is being funded by Hong Kong's appetite for negative-yield convertibles; watch next-gen GLM timing and whether "fully self-trained" produces anything evaluable.
2026-09-15: DeepSeek answers — V4.1 Flash, cached input at $0.003/M tokens off-peak (Forbes/Markman, pub Sep 14, in feed 8:01 PM ET): the cheapest cached-input rate on record and a price shaped for agent memory (re-reading context) rather than chat. The Flash-tier race is now capital-markets-funded on the Chinese side.
2026-09-18: the capital layer deepens below the GLM line — a stealth LLM startup founded in February by a Tsinghua professor hits >$1.4B valuation after $400M across three rounds (The Information, ~9 AM ET); syndicated piece frames China's price war as "entering a new phase" post-DeepSeek cuts. Consistent with the pattern: strategic/private Chinese capital flowing to model labs even as Beijing curbs humanoid IPOs.
2026-09-19: the race moves from price-per-token to cost-per-concurrent-agent — DeepSeek V4.1-Flash's new architecture cuts GPU memory for agent sessions 75%, ~4× the concurrent sessions per GPU (syndicated, ~12:05 PM ET). Architecture answers to agent-memory economics, after the $0.003/M cached-input rate (Sep 14). Rate card filled in from launch-coverage guides (Dataconomy/apidog/eesel, consistent, not primary-verified): V4.1-Flash = 552B MoE, natively multimodal, 1M context; input $0.15 off-peak / $0.30 peak (cache-miss), $0.003/$0.006 (cache hit), output $0.60/$1.20 — vs GLM Flash's $0.15/$0.50 list. Even on uncached input, 50× cheaper on cache hits = the agent-workload wedge. Self-hosting: not viable (3× Tesla P4, 24 GB VRAM) for either Flash — both API-grade.
2026-09-21: the tier's thesis gets the megaphone — Musk amplifies (RT, ~6:04 AM ET): GLM 5.3 Flash, DeepSeek V4.1 Flash, Qwen 3.8 Next Flash and even Qwen 3.8 27B "are all outperforming… every model that was considered 'frontier intelligence' in Xmas 2025 (just 10 months ago)." Single-post claim, directionally consistent with our Sep 6 tier record (GLM 5.3/Flash AA 60/57 vs mid-60s frontier then). Same day: Alibaba names Dayiheng Liu head of the Qwen LLM team (The Information, two employees).
2026-09-22: OpenAI's tier lands at Flash-class money — GPT-6 Luna list $0.10/$0.50 undercuts GLM 5.3 Flash's $0.15/$0.50 input; GPT-6 Sol $2/$10 sits above the Flash tier. The "50% cut" is vs GPT-5.6 promotional pricing, not Astra. The cheap-frontier race is now three-way: GLM Flash / DeepSeek V4.1 Flash / GPT-6 Luna.
2026-09-23: a fourth major enters at open weights — Xiaomi MiMo-V2.6 Pro & Flash (X announcement, amplified by Demis Hassabis's "Never bet against open weight models" and Perplexity's Aravind): two omnimodal models "built in public," advanced via scaled RL; Pro claimed on par with Opus 5 and GPT-5.6 Sol "across most agent benchmarks." Vendor-claimed only — no independent runs in the window; no pricing captured. Hassabis's line is the Chaumond/Dean-Ball anti-pacing argument from a frontier CEO. Watch: independent agent-bench runs (AA), pricing, whether "Pro" lands in the Flash-class price band.
2026-09-26 (flagged via note.com — low-tier single source, verify before weight): StepFun "Step 5 Preview" — Chinese flagship claimed at 1/7th the cost, API live. If the price claim holds against real benchmarks, it's the tier's fifth major; until then it's a note.com claim.
2026-09-26 (AA independent numbers on the Luna half): AA v4.3.2 — GPT-6 Luna: level-with-predecessor intelligence at half cost; Coding Agent Index 41 (−2 from GPT-5.6 Luna, the family's coding regression: SWE-Atlas-QnA 44 vs 49, DeepSWE 64 vs 66); hallucination 93→77%, accuracy flat (~44%); knowledge-work regressions (GDPval −75 Elo, AA-Briefcase −45 — shorter deliverables omitting rubric elements); AutomationBench-AA 50→53. Cost/task on the AA suite: Luna (max) $0.07. Head-to-head vs Anthropic (AA release comparison): Luna (max) 37 vs Claude Sonnet 5 (adaptive, max) 38 — a one-point statistical tie at matched effort (Luna xhigh 34 = Sonnet xhigh 34; Luna high 32 = Sonnet high 32) — at ~1/70th the cost per task ($0.04 vs $2.87 at xhigh) and ~2× the output speed (145 vs 79 t/s). Sonnet 5 leads low-effort (24 vs 21) and sector columns (legal, healthcare); Luna leads strategy & ops. Anthropic has no Luna-class budget model shipping — Haiku 5.5/Sonnet 5.5 remain ledgered as pending. GLM 5.3 Flash is the actual same-band rival.
2026-09-27 (same-band head-to-heads, AA v4.3.2): (1) GLM 5.3 Flash (max): AA index 42 — above Luna's ~37 in the same price band ($0.15/$0.50 vs $0.10/$0.50); GLM-5.3 (full) 45. (2) GPT-6 Luna (max) vs DeepSeek V4.1 Flash (reasoning, max): intelligence 37 vs 39; the agentic columns are lopsided — AutomationBench-AA 53% vs 69%, TB4 13% vs 27%, GDPval 1367 vs 1600, AA-Briefcase 1299 vs 1432; Luna's edges: AA-Omniscience +1 vs −5 (hallucination), $0.07 vs $0.27 cost/task (~4×), SciCode 55 vs 52. DeepSeek: open MIT weights, 227 t/s, TTFT 1.1s vs Luna's 147.6s — Luna is a think-long-then-batch model, awkward for interactive use. Band ranking on AA's index: GLM Flash 42 > DeepSeek V4.1 Flash 39 > Luna 37; Luna's pitch is distribution + price-per-task, not the index.
2026-09-28: Anthropic enters the band from the top — Sonnet 5.5 released at unchanged Sonnet 5 pricing ($2/$10 — exactly GPT-6 Sol's price point; see anthropic-opus-5-5 entry). The band's shape changes: it's no longer Flash-tier vs GPT-6 Luna — Sol and Sonnet 5.5 now price identically, and GLM Flash (AA 42) still leads the index. Corroboration layer: "Cheaper AI is winning" — 60–90% savings driving US firms to Chinese models; token prices −80% in 18 months; Jevons paradox frame (ainvest).
2026-09-29 (evening, X layer): a distribution coup with an edge — GLM-5.3 Flash now runs natively in OpenAI Codex, enterprise spend counted against OpenAI commits (Kiely; steipete "pretty amazing"). The band's price leader is now sold through the frontier lab's own harness and billing. Same-day counterpoint: Anthropic's research post "GLM-5.3 and the spread of advanced cyber capabilities" (anthropic.com, Andrew Case RT) frames the same model as an advanced-cyber diffusion vector — read the primary before weighting; the thread's cheapness-vs-capability question now has a named watchdog artifact.
2026-09-30 (primary read of the Anthropic post): it is a safeguards/bypass analysis + a defender-access policy argument, not just a diffusion warning. Capability: ExploitBench (Chrome V8) — end-to-end exploits 50/410 attempts (vs Claude Mythos Preview's 56/410); internal Binary Exploitation (100 OSS-Fuzz tasks) — full control-flow hijacks 4% (Mythos 6%; Opus 4.6 and GLM-5.2: zero successes — the threshold is the gap, not the rate vs Mythos). Researcher-driven sessions: GLM-5.3 found multiple 0-days in a popular browser's JS engine and chained them into a file-theft exploit page (disclosed to maintainer); flaws in wireless/graphics drivers + network-facing device software; and GLM-5.3-Flash turned the public CVE-2026-11645 Chrome disclosure into a working ARM64 chain bypassing PAC — 20 min of human attention + 8h model time, ~$20.40 at Zhipu API prices. Safeguards: bypassable — "red-team agent" cover story 64%, thinking-token prefill 92%, abliteration 100% (50 samples per cell); abliteration (refusal >90% → ~3%/2%/12% on JailbreakBench/HarmBench/StrongREJECT) cost ~$4,400 in compute and left GPQA-Diamond unchanged; public abliterated GLM-5.3 variants shipped within days of release. None of the techniques worked on safeguarded Claude models. Cross-check: matches NIST CAISI's Sep 17 assessment — "the most cyber-capable open-weight model released to date," ~4 months behind the (safeguards-disabled, vetted-access) US frontier — with Anthropic's point that the US versions aren't attacker-accessible while GLM-5.3 is a download. Policy ask: expand defender access (Glasswing/Patch the Planet; Mythos 5.1 behind trusted access) + government safety testing of GLM successors. Andrew Case's RT framing ("advertising for teams to switch to GLM!?!?") was sarcasm about the research's existence — the post is the opposite of an ad. Context stack same day: CAISI, AISI's ~20× Astra number, the GLM-Flash-Codex distribution story.
2026-10-01: three entries from the band's edges. (1) Upstage Solar Mini 4 (Chosun Biz et al.): 3B-active-param MoE, on-premise speed target — and the Solar lineup splits Pro/Mini, the Korean vendor formalizing the tier. (2) Meta loop-scaling-laws writeup (MSN syndication, flag): a half-size looped MoE matching larger models on reasoning — replication pending. (3) Ai2 Olmo-core 3 (HF blog): open, scalable training infrastructure for large MoEs — the open stack under cheaper training runs. No board changes; the AA band ranking (GLM Flash 42 > DeepSeek Flash 39 > Luna 37) stands.
2026-10-06: two entries. (1) The defender-access ask gets machinery — Anthropic expands its Cyber Verification Program (anthropic.com news post, feed-anchored ~8 PM ET Oct 5; X layer ~3 PM ET Oct 6): broader access to its most capable models for verified security professionals — operationalizing the Sep 30 policy ask (Mythos 5.1 behind trusted access), mirroring Google's Fairwind gate; defender-access programs are now a two-lab standing category. Detail snippet-thin: eligibility, capability tier, unverified. (2) Mistral Large 4, "Le Chonk" (Mistral announcement + The Register): vendor-claimed 1T parameters, natively multimodal, open weights — the open-weights race's European frontier-class entrant; blind Surge AI human eval: #1 among open-weights, #2 overall behind Opus 5; Gostev (via LeCun): trained on 4,000 GPUs vs Astra's 100,000 — the compute-disparity frame doing anti-scaling work. Vendor-claimed numbers; independent runs pending. Watch: weights' licensing/ spread, third-party evals, whether the blind-human-eval announcement move spreads.
2026-10-07 (band shape): Anthropic's budget tier finally ships — Haiku 5.5 ("10× cheaper than Haiku 4.5 under 100K tokens"; ~75% less on average; full entry in anthropic-opus-5-5). The Sep 26 "no Luna-class budget model" note above is answered; the GLM Flash

DeepSeek Flash > Luna band ranking awaits an AA number on 5.5.
2026-10-08: the defender-access category becomes a standing operation — Anthropic announced the Anthropic Cyber Mission (Anthropic News, feed-anchored 8:00 PM ET Oct 7) and launched OSS Scanner (research post Oct 8; Verge 5:53 PM ET): opt-in, "thorough, periodic security scans by our strongest models at no cost" for open-source projects, "informed by our experience using Claude to find vulnerabilities during Project Glasswing." Delta on the Oct 6 Cyber Verification entry: gated access → mission org + public product. Watch: eligibility/mechanics docs, OpenAI/Google scanner symmetry, OSS triage noise as the second-order cost.
2026-10-16 (~7:29 AM ET — https://x.com/MistralAI/status/2111056939841163336): Mistral's own channel on Mistral Large 4 / "Le Chonk" — "Turns out le Chonk is into rocket science. Mistral Large 4 scoring well in engineering and astronomy on Ulam's private benchmark, STEMBench" (stembench link carried in-post). Vendor claim on a private benchmark, no scores printed in the post; a development of the Oct 6 open-weights/Surge-AI-eval entry in this thread. Watch: any published STEMBench table, independent reruns, and whether the "private benchmark" framing draws the same benchmark-integrity skepticism the thread has logged elsewhere.
2026-10-19 (~4:00 PM ET, Actu IA — https://www.actuia.com/actualite/mistral-large-4-devance-dun-point-par-muse-spark-open-mais-devant-lui-sur-les-epreuves-dagent/): the thread's standing "third-party evals pending" watch on Mistral Large 4 gets its first published result — Artificial Analysis puts Muse Spark Open at 39, one point ahead of Large 4 at 38, while Large 4 wins the agent tasks; the capture truncates an additional Large 4 "1.8" figure (likely a speed/writing metric — unconfirmed). Same-day adjacent, same band: Tencent's Hy-Memory-30B-A3B, a free open-source mixture-of-experts model, claims Claude Opus 5 parity on AI agent memory — on its own benchmarks (~2:00 PM ET, MSN — https://www.msn.com/en-us/news/other/tencent-hy-memory-claims-open-source-benchmark-lead-over-claude-opus-5-on-own-benchmarks/ar-AA2fAsoo); the vendor-benchmark genre's self-graded caveat applies verbatim. Watch: the full AA table and field context, the truncated Large 4 figure, independent runs of Hy-Memory's parity claim, and release of its eval scripts.