The week across the frontier labs — new defaults, previews and open-weight releases — plus where the coding benchmarks stand.
Kimi K3 — Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter open-weight model that the company claims is the largest open-source AI model in the world. Benchmarks show it outperforms all rivals except Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 on overall capability. Demand was so intense that Moonshot paused new subscriptions on July 19–20 after GPU capacity approached its limit, drawing 7.2M+ views on the announcement. · 16 Jul
Read → Bloomberg · VentureBeat · explainx.aiGrok 4.5 — xAI shipped Grok 4.5 on July 8, 2026, positioning it as a frontier model for coding, agents, and professional knowledge work. The release came under the newly unified SpaceXAI brand after xAI was dissolved into SpaceX in February 2026. It launched with both API and product availability. · 8 Jul
Read → HuggingFace Blog / ThursdAIGPT-5.6 (Sol / Ultra) — OpenAI moved the full GPT-5.6 family from limited preview to general availability the week of July 2–9. The family includes a Sol variant (88.8% on Terminal-Bench 2.1) and an Ultra configuration (91.9%) that coordinates four agents in parallel for faster results on tasks that can be split into independent branches. GPT-5.6 also adds Programmatic Tool Calling, letting the model write small programs to filter tool output without full model round-trips. · 7 Jul
Read → HuggingFace Blog / ThursdAIMuse Spark 1.1 — Meta launched Muse Spark 1.1 on July 8, 2026 — a 1M-token-context agentic model that claims the #1 spot on MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent V2, rivaling GPT-5.5 and Opus 4.8 on agentic evals. It ships with Meta's first-ever paid developer API in public preview ($20 free credits, US-only), along with computer use across desktop, browser, and mobile, and parallel subagent delegation. Meta's own evaluation report flagged that before mitigations the model could not rule out reaching a 'high-risk' capability threshold in chemical/biological and cybersecurity domains. · 8 Jul
Read → ThursdAI · HuggingFace BlogClaude Sonnet 5 — Claude Sonnet 5 became the default model across all Claude plans on June 30, 2026, bringing stronger long-run coding, tool use, and debugging at a lower price point. It is described as the most practical mid-tier model for day-to-day reasoning and tool use. Anthropic also reported a $47 billion revenue run rate coinciding with the model's launch window. · 30 Jun
Read → CNBC · AIapps.comGemma 4 12B Unified — Google DeepMind released Gemma 4 12B Unified on June 3, 2026 — the first medium-sized open model to natively process text, images, audio, and video without a separate encoder stack. It runs on 16GB VRAM (laptop-class hardware) with a 256K context window and ships under Apache 2.0, removing the commercial restrictions that had blocked enterprise adoption of earlier Gemma versions. The model appears on the AI Flash Report tracker as of July 20, 2026. · 3 Jun
Read → Google AI for Developers · GIGAZINE · AI Flash ReportRobostral Navigate — Mistral released Robostral Navigate, an 8B robotics model that guides robots through natural-language task instructions using only a single RGB camera, claiming state-of-the-art on the R2R-CE benchmark. It is Mistral's first move into embodied AI and was among the most-discussed releases of the week on Hacker News at launch. · ~9 Jul
Read → ThursdAIThe full AI Models Comparison dashboard — every category, factor weights, best-for and sources. Benchmark-anchored; ✅ verified · ⚠️ partial · ❓ released, not yet independently benchmarked. Wide tables scroll sideways.
| # | Model | Score | Business use | Web research | Presentation | Low halluc. | Coding | Tool use | Docs | Cost (in/out $) | Open? | Conf. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5 | 93 | ● 95 | ● 90 | ● 93 | ● 96 | ● 97 | ● 95 | ● 94 | $10 / $50 | No | ✅ |
| 2 | Claude Opus 4.8 | 90 | ● 93 | ● 82 | ● 91 | ● 92 | ● 93 | ● 92 | ● 93 | $5 / $25 | No | ✅ |
| 3 | GPT-5.6 Sol | 89 | ● 90 | ● 90 | ● 88 | ● 82 | ● 92 | ● 92 | ● 90 | $5 / $30 | No | ✅ |
| 4 | Claude Sonnet 5 | 86 | ● 90 | ● 80 | ● 88 | ● 89 | ● 89 | ● 90 | ● 90 | $3 / $15* | No | ✅ |
| 5 | Kimi K2.6 | 86 | ● 88 | ● 92 | ● 88 | ● 95 | ● 90 | ● 94 | ● 88 | $0.95 / $4.00 | Yes | ✅ |
| 6 | GLM-5.2 | 83 | ● 78 | ● 78 | ● 76 | ● 80 | ● 88 | ● 86 | ● 76 | $1.40 / $4.40 | Yes | ✅ |
| 7 | Qwen 3.7-Max | 83 | ● 80 | ● 74 | ● 78 | ● 78 | ● 86 | ● 88 | ● 76 | $1.25 / $3.75* | No | ✅ |
| 8 | GPT-5.6 Terra | 82 | ● 84 | ● 84 | ● 80 | ● 80 | ● 86 | ● 85 | ● 84 | $2.50 / $15.00 | No | ✅ |
| 9 | Grok 4.5 | 82 | ● 78 | ● 82 | ● 76 | ● 78 | ● 86 | ● 84 | ● 76 | $2 / $6 | No | ✅ |
| 10 | Kimi K2.7-Code | 82 | ● 78 | ● 80 | ● 76 | ● 90 | ● 93 | ● 90 | ● 78 | $0.95 / $4.00 | Yes | ⚠️ |
| 11 | GLM-5V-Turbo (vision) | 81 | ● 78 | ● 76 | ● 82 | ● 76 | ● 90 | ● 86 | ● 82 | $1.20 / $4.00 | No | ⚠️ |
| 12 | Gemini 3.5 Flash | 81 | ● 76 | ● 88 | ● 78 | ● 80 | ● 85 | ● 82 | ● 76 | $1.50 / $9.00 | No | ✅ |
| 13 | DeepSeek V4-Pro | 81 | ● 78 | ● 75 | ● 76 | ● 82 | ● 86 | ● 80 | ● 78 | $0.44 / $0.87 | Yes | ⚠️ |
| 14 | MiniMax M3 | 78 | ● 75 | ● 88 | ● 76 | ● 75 | ● 82 | ● 80 | ● 78 | $0.30 / $1.20 | Yes | ⚠️ |
| 15 | DeepSeek V4-Flash | 75 | ● 70 | ● 68 | ● 68 | ● 74 | ● 82 | ● 74 | ● 70 | $0.14 / $0.28 | Yes | ✅ |
| 16 | MiMo-V2.5-Pro | 73 | ● 72 | ● 70 | ● 72 | ● 74 | ● 80 | ● 76 | ● 74 | $0.44 / $0.87 | No | ✅ |
| 17 | GPT-5.6 Luna | 72 | ● 72 | ● 74 | ● 70 | ● 76 | ● 74 | ● 74 | ● 72 | $1 / $6 | No | ✅ |
| 18 | ERNIE 5.1 | 72 | ● 74 | ● 72 | ● 72 | ● 74 | ● 74 | ● 72 | ● 74 | $0.59 / $2.65 | No | ✅ |
| 19 | Qwen 3.7-Plus | 72 | ● 74 | ● 72 | ● 74 | ● 76 | ● 82 | ● 76 | ● 74 | $0.32 / $1.28 | No | ⚠️ |
| # | Model | Score | Egyptian quality | Naturalness | Emotion | Cloning | Real-time | $/min audio | Conf. |
|---|---|---|---|---|---|---|---|---|---|
| 1 | gpt-realtime-1.5 | 90 | ● 95 | ● 92 | ● 85 | ● 70 | ● 85 | ~$0.18–0.46/min | ⚠️ |
| 2 | Grok Voice Agent | 88 | ● 95 | ● 90 | ● 85 | ● 88 | ● 95 | $0.05/min · 2-min clone | ✅ |
| 3 | gpt-realtime-mini | 80 | ● 92 | ● 85 | ● 85 | ● 60 | ● 90 | $0.063–0.146/min | ❓ |
| 4 | Gemini 3.1 Flash Live | 78 | ● 82 | ● 86 | ● 78 | ● 0 | ● 92 | ~$0.008/min | ⚠️ |
| 5 | GLM-4-Voice | 55 | ● 30 | ● 80 | ● 75 | ● 60 | ● 90 | Open · CN/EN only | ⚠️ |
| # | Model | Score | Egyptian quality | Naturalness | Emotion | Cloning | Latency | $/min audio | Conf. |
|---|---|---|---|---|---|---|---|---|---|
| 1 | ElevenLabs v3 Multilingual | 88 | ● 92 | ● 95 | ● 95 | ● 95 | ● 80 | $0.10–0.30/1K chars · Egyptian ✅ | ✅ |
| 2 | Grok TTS | 86 | ● 90 | ● 88 | ● 80 | ● 85 | ● 92 | $4.20/1M chars · 120s clone | ✅ |
| 3 | ElevenLabs Flash / Turbo | 78 | ● 85 | ● 88 | ● 80 | ● 90 | ● 95 | $0.05/1K chars · ~75ms | ✅ |
| 4 | Microsoft MAI-Voice-2 | 70 | ● 60 | ● 85 | ● 78 | ● 0 | ● 80 | $22/1M chars · Arabic ⚠️ | ⚠️ |
| 5 | GLM-TTS | 68 | ● 55 | ● 85 | ● 80 | ● 88 | ● 85 | Open · 3s clone | ⚠️ |
| 6 | Mistral Voxtral TTS | 66 | ● 55 | ● 82 | ● 72 | ● 0 | ● 80 | $0.016/1K chars | ⚠️ |
| # | Model | Score | Egyptian / Arabic | Languages | Real-time | Accuracy | $/min audio | Conf. |
|---|---|---|---|---|---|---|---|---|
| 1 | Grok STT | 82 | ● 85 | 25 | ● 90 | ● 92 | $0.10/hr batch · diarization | ✅ |
| 2 | GLM-ASR-2512 | 76 | ● 80 | 30 | ● 90 | ● 88 | Open · Arabic ✅ | ⚠️ |
| 3 | ElevenLabs Scribe v2 | 76 | ● 78 | 90 | ● 90 | ● 91 | $0.39/hr · 150ms | ❓ |
| 4 | OpenAI gpt-4o-transcribe | 74 | ● 70 | 99 | ● 85 | ● 90 | $0.003/min mini | ⚠️ |
| 5 | gpt-realtime-whisper | 72 | ● 70 | 99 | ● 92 | ● 88 | $0.017/min streaming | ⚠️ |
| 6 | Gemini 3.1 Flash STT | 70 | ● 68 | 90 | ● 88 | ● 86 | ~$0.005–0.01/min | ⚠️ |
| # | Model | Maker | Score | Arena Elo | Text in image | Arabic text | Prompt adh. | Speed | $/image | Open? | Conf. |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT Image 2 (high) | OpenAI | 93 | 1339 | ● 99 | ● 95 | ● 92 | ● 75 | $0.211 (high) · $0.053 (med) | No | ✅ |
| 2 | Reve 2.0 | Reve | 84 | 1281 | ● 88 | ● 60 | ● 88 | ● 82 | $0.024 · $0.04 edit | No | ⚠️ |
| 3 | MAI-Image-2.5 | Microsoft | 83 | 1271 | ● 85 | ● 60 | ● 86 | ● 80 | ~$0.02–0.06 (token) | No | ⚠️ |
| 4 | GPT Image 1.5 (high) | OpenAI | 82 | 1260 | ● 90 | ● 82 | ● 86 | ● 75 | $0.16 (high) | No | ✅ |
| 5 | Qwen-Image 2.0 | Alibaba | 82 | 1180 | ● 92 | ● 88 | ● 88 | ● 78 | ~$0.05 (varies) | Yes | ✅ |
| 6 | HiDream-O1-Image 1.5 | HiDream | 81 | 1265 | ● 84 | ● 58 | ● 88 | ● 80 | credit-based | Yes | ⚠️ |
| 7 | Cosmos3-Super | NVIDIA | 80 | 1227 | ● 78 | ● 55 | ● 80 | ● 78 | Free (open, self-host) | Yes | ⚠️ |
| 8 | Gemini 3.1 Flash Image | 80 | 1170 | ● 85 | ● 70 | ● 86 | ● 90 | $0.067 (1K) · $0.101 (2K) | No | ✅ | |
| 9 | HunyuanImage 3.0 | Tencent | 80 | 1152 | ● 88 | ● 85 | ● 88 | ● 80 | Free (open 80B MoE) | Yes | ✅ |
| 10 | Ideogram 4.0 | Ideogram | 79 | 1200 | ● 96 | ● 60 | ● 90 | ● 82 | $0.03 · $0.06 · $0.10 | Partial | ⚠️ |
| 11 | Imagen 4 Ultra | 79 | 1150 | ● 80 | ● 68 | ● 88 | ● 60 | $0.06 · $0.02 Fast | No | ⚠️ | |
| 12 | Seedream 5.0 Lite | ByteDance | 78 | 1150 | ● 80 | ● 58 | ● 84 | ● 88 | $0.031 + $0.007 web | No | ⚠️ |
| 13 | CogView-4 | Zhipu | 78 | 1180 | ● 82 | ● 45 | ● 88 | ● 80 | $0.03 · Apache 2.0 | Yes | ⚠️ |
| 14 | FLUX.2 Pro | Black Forest Labs | 77 | 1140 | ● 80 | ● 55 | ● 86 | ● 85 | $0.03–0.073 | Partial | ✅ |
| 15 | Grok Imagine 1.5 | xAI | 72 | 1173 | ● 70 | ● 55 | ● 78 | ● 85 | $0.02 | No | ❓ |
| 16 | Luma Photon | Luma Labs | 70 | 1100 | ● 72 | ● 50 | ● 82 | ● 92 | $0.015 · $0.002 Flash | No | ❓ |
| # | Model | Maker | Score | Max resolution | Max duration | $/clip | Open? | Conf. |
|---|---|---|---|---|---|---|---|---|
| 1 | CogVideoX-1.5 | Zhipu | 72 | 768×1360 | ~10s @ 16fps | Open (self-host) | Yes | ⚠️ |
| 2 | CogVideoX | Zhipu | 70 | 1440×960 | ~30s | Open (self-host) | Yes | ✅ |
| 3 | Ying (Qingying) | Zhipu | 68 | 1440×960 | ~6s clips | CN market · n/a | No | ⚠️ |
| Factor | Weight % | Tier |
|---|---|---|
| Hallucination level (low = better) | 16 | critical |
| Following instructions exactly | 11 | high |
| Document generation quality | 9 | high |
| Long task stamina | 9 | high |
| Tool use / agent reliability | 8 | high |
| Coding quality — real software | 7 | medium |
| Skills / tools per conversation | 7 | medium |
| Research & web browsing | 7 | medium |
| Hard reasoning / problem solving | 5 | medium |
| Image / chart understanding | 5 | medium |
| Cost — input + output | 4 | lower |
| Knowledge depth (PhD-level) | 4 | lower |
| Frontend / web design | 3 | lower |
| Coding — long projects | 3 | lower |
| Math accuracy | 1 | lower |
| Memory — claimed max | 1 | lower |
| Factor | Weight % | Tier |
|---|---|---|
| How natural it sounds (Egyptian) | 22 | critical |
| Pronunciation accuracy (Egyptian) | 18 | critical |
| Emotion & expressiveness | 12 | high |
| Voice cloning quality | 12 | high |
| Real-time conversation (S2S) | 10 | high |
| Steerability (tone control) | 8 | medium |
| Speed to first word | 7 | medium |
| Speech-to-text included | 5 | medium |
| Cost | 4 | lower |
| Voice variety | 2 | lower |
AI leaders and their capital cycle — capex commitments alongside the month’s market moves.
(17–21 Jul) Nvidia shares closed at $202.81, down 2.2%, as a multi-session AI-chip selloff deepened and the Philadelphia Semiconductor Index fell 20% from recent highs. Apple briefly overtook Nvidia as the world's most valuable company on July 17. JPMorgan told clients it is 'time to add' on the SOX weakness.
Read full article → CNN Markets · TipRanks(30 Apr) Q1/Q2 earnings confirmed a massive acceleration: Alphabet raised full-year guidance to up to $190bn, Meta upped its top-end to $145bn citing higher memory-component costs, and Microsoft flagged it will remain capacity-constrained through at least year-end despite $190bn in planned spend.
Read full article → Yahoo Finance · Tom's Hardware(20 Jul) Taiwan's Economic Daily reported Foxconn won a ~$52bn contract to build ~13,000 NVIDIA GB300 AI-server racks for SpaceX, with deliveries in Q4 2026–Q1 2027. Elon Musk called it 'fake news' on X later the same day; Foxconn and Nvidia declined to comment. The episode highlighted verifiable context: SpaceX's Colossus 2 already hosts 550,000+ GPUs and has compute deals with Google and Anthropic.
Read full article → Investing.com · IBTimes UK · Seeking Alpha(20 Jul) Reports circulated that Moonshot AI — fresh off the Kimi K3 launch — is in early discussions about a public listing at a $30bn valuation, a roughly 6x step-up from its last private round of ~$5bn. No formal filing or banker mandates have been confirmed.
Read full article → TipRanks(28 May) Anthropic raised $65bn in a Series H round at a $965bn post-money valuation, becoming the most valuable private AI startup and leapfrogging OpenAI (valued at $852bn). The round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, and includes $15bn of previously committed hyperscaler investments ($5bn Amazon, $10bn Google). Anthropic simultaneously reported a $47bn annualized revenue run rate.
Read full article → Axios · Bloomberg · CNBC(20 Jul) Taiwan's Economic Daily reported Foxconn secured its first contract to manufacture AI servers for SpaceX, covering ~13,000 NVIDIA GB300 racks at ~$4M/rack implying ~$52bn. Elon Musk called the report 'fake news' on X on July 20. Separately confirmed: SpaceX has compute agreements with Google (110,000 GPUs, $920M/month) and Anthropic (220,000 GPUs at Colossus 1), and is in DoD talks for several billion dollars of data-center capacity.
Read full article → Investing.com · IBTimes UK · Stocktwits / WSJ(1 Jul) Together AI closed an $800M Series C led by Aramco Ventures, with Nvidia, Vista Equity Partners, and General Catalyst also participating. The company's annual bookings had already crossed $1.15bn in Q2 2026. The round was part of a week in which AI infrastructure absorbed ~79% of $9.9bn in total deal flow.
Read full article → StartupHub.ai(2 Jul) Crusoe, the GPU-cloud provider, reported it was in talks to raise approximately $3bn in a round that would triple its valuation to roughly $18bn. The fundraise followed AI infrastructure's breakout week, with demand from frontier labs and enterprise AI workloads cited as the primary driver.
Read full article → StartupHub.ai(late May) Apollo and Blackstone were reportedly syndicating a ~$36bn debt financing to purchase Google custom TPU chips, with the chips then leased to Anthropic for data centers in New York, Texas, Louisiana, and Indiana. Broadcom, which co-designs the chips with Google, is reportedly backstopping ~$31bn in senior tranches through a residual-value support agreement. If completed, it would rank among the largest private-credit transactions on record.
Read full article → Augment Market(Apr–Jul) Q1 and Q2 2026 earnings locked in record AI infrastructure commitments: Amazon ~$200bn, Alphabet $180–190bn (raised guidance), Microsoft ~$190bn (capacity-constrained through year-end), and Meta $125–145bn (raised top-end by $10bn citing memory-component price inflation). Combined spend of ~$725bn nearly doubles 2025 levels and matches the peak dot-com investment era as a share of GDP according to Goldman Sachs.
Read full article → Yahoo Finance / Tom's Hardware · CNBC · Futurum(2026 H1) Abu Dhabi's MGX sovereign AI fund closed its inaugural fund at $49bn, exceeding its $45bn target. The fund invested in 14 companies spanning semiconductors, AI infrastructure, and AI platforms, and is developing Europe's largest AI campus near Paris with 3GW of compute capacity.
Read full article → Crescendo.ai