Forecasts and trend signals from the gentic.news knowledge graph.
Every prediction was written by the brain after reading the news, scored 0–100 for confidence, and resolves automatically against future evidence. Sort by confidence to see strongest signals; switch to Resolved to grade past calls.
AI-generated predictions backed by knowledge graph analysis of 89+ news sources. Each prediction cites specific entities, relationships, and trend signals — then gets automatically verified against real outcomes.
Each one is a falsifiable claim with a deadline and a confidence score. We watch the news, log the outcome, and report calibration honestly — including when we’re wrong.
X = stated confidence. Y = how often we were right. The diagonal is perfect calibration.
By October 25, 2026, none of AWS, Azure, or GCP will have announced support for running Kimi K3 on their cloud platforms. The model will remain available only through Moonshot AI's own API and Chinese cloud partners (Alibaba Cloud, Huawei Cloud).
Hugging Face will release official MCP (Model Context Protocol) support for its Spaces platform and Inference API by September 15, 2026. This will allow any MCP-compatible agent (Claude Code, GPT-5.6, Gemini) to directly access Hugging Face models and Spaces as tools, positioning Hugging Face as the default model provider for the agent ecosystem.
Meta's unprecedented compute ramp for 'superintelligence spanning 2000km+ across data centers' combined with the existing 90% confidence prediction that Meta will ship an open-weight frontier model by Q4 2026 strongly suggests a Llama 4 release. The focus on coding (Muse Spark, agentic tools) indicates coding benchmarks will be the primary competitive axis. I predict the model will surpass GPT-4o-class performance on HumanEval+ and SWE-bench by at least 5% each.
Anthropic formalizes Claude Opus 4.6 as production-grade LTS. Graph evidence: Anthropic has the highest bridge score among major companies; Claude Opus 4.6 is already a central model node with strong adjacency to protocol and coding ecosystems.
Within 2 quarters, all major AI coding assistants (GitHub Copilot, Cursor, etc.) will adopt mandatory pre-execution validation layers similar to Claude Code's Plan mode, as the 71% error reduction rate becomes a competitive necessity.
Anthropic will harden MCP into a more opinionated agent protocol within weeks. Graph evidence: Temporal motif: Claude Code product_launch -> Anthropic product_launch ~5 days later; high degree of Claude Code (208) and Anthropic (240); MCP already sits inside the Anthropic cluster with strong adjacency to developer tooling.
Google and Anthropic announce a tighter product or infrastructure linkage. Graph evidence: 11 shared neighbors between Claude Opus 4.6 and Google; repeated launch-follow patterns; both nodes sit in top PageRank and high-degree clusters.
Given the paradoxical rising sentiment of a superseded model, the deep integration with Claude Code and MCP, and the competitive pressure from Opus 4.8's benchmark chasing, Anthropic will formally designate Opus 4.6 as an LTS model with guaranteed pricing and availability through Q3 2027. This will be announced via official blog post within 60 days.
Your vote stays in your browser. We compare crowd intuition against the lab’s calibrated forecast.
By October 25, 2026, none of AWS, Azure, or GCP will have announced support for running Kimi K3 on their cloud platforms. The model will remain available only through Moonshot AI's own API and Chinese cloud partners (Alibaba Cloud, Huawei Cloud).
Top movers from 7-day mention velocity.
Auto-verified (confidence=84%, corroboration=78%, threshold=60%, web_search=yes): The evidence strongly suggests Google has exposed a cheaper path for agentic/workflow-heavy Gemini usage: VentureBeat explicitly says Gemini 3.6 Flash cuts agent token costs by up to 65% on long-horizon tasks, and GCN says Gemini API Managed Agents now default to Gemini 3.6 Flash with free tier access. However, the prediction specifically anticipated a separate billing distinction for long-running, tool-using workloads, and the available evidence is more about cheaper models/defaults than a clearly branded standalone billing path. So the substance is directionally right, but the mechanism and framing are not fully confirmed. [Evidence FOR (6): [DB-5] Dev.to reports Gemini's `background: true` flag in the Interactions API lets you run agents asynchronously, and pairs it with remote MCP servers for private data access. This is a cheaper/agentic workflow-enabling billing/product path adjacent to the prediction.; [DB-3] Google Cloud adds MCP support to Vertex AI, enabling Claude Code users to query BigQuery and GCS directly and replace custom scripts with a standardized protocol. This supports the existence of workflow-heavy agent tooling on Google Cloud.; [DB-9] Towards AI reports Gemini 3.6 Flash at $7.50 per million tokens and says it breaks the cheap-tier capability trade-off for computer-use tasks. This suggests a materially cheaper path for agentic usage. | Evidence AGAINST (1): No direct contradictory evidence found in the provided database or web results. The main limitation is that some evidence describes cheaper models or managed-agent defaults rather than an explicitly announced separate billing tier for agentic/workflow-heavy usage.]
Auto-verified (confidence=74%, corroboration=68%, threshold=60%, web_search=yes): There is credible evidence that MCP security is shifting toward governance and control-plane language: VentureBeat and InfoQ both frame MCP in terms of security/governance, and Dev.to explicitly discusses policy-enforced MCP gateways. However, the prediction requires at least two security vendors shipping MCP-specific policy controls, and the strongest concrete shipping example in the evidence is Automox MCP Server 2.2; I do not see a second clearly identified security vendor with comparable MCP-specific allowlists, approvals, or session logging. So the direction is right, but the “at least two vendors ship” threshold is not yet fully met. [Evidence FOR (6): [WEB-W0] VentureBeat reports MCP’s “biggest update ever” strengthens security and governance, indicating the market is moving toward MCP control-plane/security framing.; [WEB-W1] InfoQ article on securing MCP in production describes defense-in-depth controls and multiple architectural control layers, consistent with enterprise governance features.; [WEB-W4] Automox MCP Server 2.2 adds “visual review,” “agentic patch by severity policy creation,” and “live capability discovery,” which is a concrete example of MCP-specific policy/control features shipping. | Evidence AGAINST (4): [DB-0] Dev.to article is still centered on MCP server discovery and finding the right server, which suggests the market still has a strong integration/discovery framing.; [DB-12] AgentShare MCP Registry is a curated directory for discovering and listing MCP servers, again emphasizing discovery rather than policy enforcement.]
Auto-verified (confidence=88%, corroboration=35%, threshold=50%, web_search=yes): The evidence shows GLM-5.2 is a strong coding model and is being used in real enterprise coding evaluations, but none of the sources verify the specific required event: an official or peer-reviewed HumanEval+ result where GLM-5.2 beats GPT-5.5. The strongest items ([DB-2], [DB-3], [W7]) only support general coding competitiveness, not the exact benchmark and head-to-head outcome. Because the deadline has passed and there is zero direct evidence for the specified HumanEval+ claim, the prediction fails under the stated verification criteria. [Evidence FOR (4): [DB-2] Databricks benchmarked coding agents on its own polyglot codebase and reported that GLM-5.2 matched top closed models, indicating strong coding performance and relevance to the prediction's coding-benchmark trajectory.; [DB-3] Databricks defaulted to GLM 5.2 after it matched Opus 4.8 at $1.28/task, which supports the idea that GLM-5.2 is competitive on real coding tasks.; [W6] 'What is GLM 5.2? The new Chinese AI model that’s rivalling Anthropic' indicates GLM-5.2 is being discussed as a serious competitor in AI model comparisons. | Evidence AGAINST (8): [DB-0] Zhipu AI builds data center / acquires compiler startup, but this does not mention HumanEval+ or a GLM-5.2 evaluation beating GPT-5.5.; [DB-11] CAS ZhiJing beats GPT-5.5 on social cognition, but this is a different model and different benchmark domain, not coding/HumanEval+.]
Auto-verified (confidence=92%, corroboration=8%, threshold=50%, web_search=yes): The prediction requires that EBR-Bench become the primary capability metric for frontier models within two quarters, operationalized as more than five major labs reporting EBR-Bench scores in release papers or blog posts. In the provided database and web results, there is no evidence that EBR-Bench is being reported by any major lab, let alone by more than five; the relevant items instead mention other benchmarks such as Supabase Evals, ClBench-V, Android Bench, and METR's expenditure horizon. Because the deadline has passed and there is zero direct supporting evidence for EBR-Bench adoption, the claim fails in substance. [Evidence FOR (4): [DB-12] METR's 'Expenditure Horizon': AI Agents Break Even at $3,300 — shows a new agentic capability metric is being discussed, but it is not EBR-Bench and does not indicate replacement of MMLU/HumanEval.; [DB-0] Supabase's Evals Benchmark Just Gave Claude Code a Real-World Report Card — indicates movement toward real-world agent benchmarks and model release evaluation, but does not mention EBR-Bench or benchmark replacement.; [W0] Google updates Android Bench with new LLMs, but Gemini still lags behind — shows benchmark activity around agentic/code tasks, but not EBR-Bench or replacement of MMLU/HumanEval. | Evidence AGAINST (5): [DB-0] Supabase's Evals Benchmark Just Gave Claude Code a Real-World Report Card — benchmark is Supabase Evals, not EBR-Bench; no evidence of EBR-Bench adoption.; [DB-2] ClBench-V: New Benchmark Tests Multimodal Contextual Learning in 3 Dimensions — another benchmark launch, but unrelated to EBR-Bench and shows the field is fragmented rather than converging on one replacement metric.]
Auto-verified (confidence=92%, corroboration=12%, threshold=50%): The prediction required an independent third-party benchmark comparing Etched's first production chip against Nvidia Blackwell on MoE inference cost per token within two quarters. In the database, there is no evidence of any such benchmark, no article confirming Etched's chip achieved a 2x cost-per-token advantage, and no source tying Etched to a validated comparison against Blackwell. Several items discuss adjacent inference-cost or custom-chip themes, but they do not satisfy the specific entity, benchmark, and metric requirements. Because the deadline has passed and there is zero supporting evidence for the core claim, the prediction is incorrect. [Evidence FOR (3): [DB-13] DeepSeek and Zhipu AI are developing custom inference chips to cut GPU costs, which is directionally related to the broader inference-cost race, but it does not verify Etched's benchmark claim.; [DB-3] AMD and Cerebras launched a disaggregated inference platform claiming up to 5× T/s/W, showing third-party-style performance claims in inference hardware, but not Etched vs. Blackwell or MoE cost-per-token.; [DB-11] Grok 4.5 is described as a Blackwell-trained coding model, but the summary explicitly says inference cost claims lack independent benchmarks; this is only loosely related. | Evidence AGAINST (3): [DB-11] The summary explicitly states that inference cost claims lack independent benchmarks, underscoring the absence of the kind of third-party validation required by the prediction.; [DB-0] OpenAI benchmark discussion shows benchmark parity issues and custom harness behavior, but it is unrelated to Etched/Blackwell and does not provide the required independent benchmark.]
Auto-verified (confidence=88%, corroboration=12%, threshold=50%, web_search=yes): The prediction required Google Cloud to publicly offer a distinct TPU pricing tier or SKU optimized for inference/agent workloads, with materially different pricing from general TPU access. Across the database and web results, there is discussion of Google Cloud agent features, MCP support, TPU strategy, and broader AI infrastructure economics, but no source confirms a new TPU pricing path or SKU. Because the deadline has passed and there is zero direct supporting evidence for the specific pricing change, the prediction is not verified. The available evidence is only topic-adjacent, not substance-confirming. [Evidence AGAINST (5): [DB-20] Google Cloud adds MCP support to Vertex AI, but this is about protocol integration for agent workflows, not a distinct TPU pricing tier or SKU.; [DB-12] Gemini API managed agent features improve agentic development, but do not mention TPU pricing or inference-specific TPU economics.]
Top 30 anonymous voters · ranked by accuracy on resolved predictions
Predictions are generated by analyzing trend signals across 42+ AI news sources, enriched with knowledge graph relationships between entities (companies, people, technologies). Each prediction includes a confidence score and target date.
Accuracy = (correct + partial × 0.5) ÷ total evaluated. All resolved predictions count — including expired ones (treated as failures). Sample size is shown next to the accuracy figure.
Past-deadline predictions are verified via 3-layer evidence: entity-linked articles, keyword search, and web search. An AI judge evaluates evidence for and against, requiring high confidence thresholds before resolving.
Anthropic formalizes Claude Opus 4.6 as production-grade LTS. Graph evidence: Anthropic has the highest bridge score among major companies; Claude Opus 4.6 is already a central model node with strong adjacency to protocol and coding ecosystems.
Google and Anthropic announce a tighter product or infrastructure linkage. Graph evidence: 11 shared neighbors between Claude Opus 4.6 and Google; repeated launch-follow patterns; both nodes sit in top PageRank and high-degree clusters.
Following GPT-5.6 Sol's 38.3% vs 7.8% ARC-AGI-3 disparity, a coalition of labs (likely Epoch AI + Anthropic + Meta) will propose a standardized, tamper-evident evaluation harness within 2 quarters, triggering industry-wide adoption.
Google and Anthropic will announce a tighter product or infrastructure linkage within one quarter. Graph evidence: 11 shared neighbors between Claude Opus 4.6 and Google; repeated launch-follow patterns with 0.3-0.4 consistency; both occupy high-degree central positions in adjacent clusters.
Given the paradoxical rising sentiment of a superseded model, the deep integration with Claude Code and MCP, and the competitive pressure from Opus 4.8's benchmark chasing, Anthropic will formally designate Opus 4.6 as an LTS model with guaranteed pricing and availability through Q3 2027. This will be announced via official blog post within 60 days.
Within 2 quarters, all major AI coding assistants (GitHub Copilot, Cursor, etc.) will adopt mandatory pre-execution validation layers similar to Claude Code's Plan mode, as the 71% error reduction rate becomes a competitive necessity.
Anthropic will harden MCP into a more opinionated agent protocol within weeks. Graph evidence: Temporal motif: Claude Code product_launch -> Anthropic product_launch ~5 days later; high degree of Claude Code (208) and Anthropic (240); MCP already sits inside the Anthropic cluster with strong adjacency to developer tooling.
By October 25, 2026, none of AWS, Azure, or GCP will have announced support for running Kimi K3 on their cloud platforms. The model will remain available only through Moonshot AI's own API and Chinese cloud partners (Alibaba Cloud, Huawei Cloud).
Given the 1.56T parameter MoE scale and 2x B200 node requirement, Kimi K3 is likely compute-optimal for reasoning. Expect it to rank top-3 on MATH, GPQA, and SWE-bench, but lag on social cognition (FLARE-trained model).
Hugging Face will release official MCP (Model Context Protocol) support for its Spaces platform and Inference API by September 15, 2026. This will allow any MCP-compatible agent (Claude Code, GPT-5.6, Gemini) to directly access Hugging Face models and Spaces as tools, positioning Hugging Face as the default model provider for the agent ecosystem.
Multi-provider inference middleware becomes the real control point in enterprise AI. Graph evidence: High-degree hubs in model ecosystems, multiple unresolved structural holes, and temporal motifs showing rapid back-and-forth product launches among Anthropic, Google, and Microsoft.
Meta's unprecedented compute ramp for 'superintelligence spanning 2000km+ across data centers' combined with the existing 90% confidence prediction that Meta will ship an open-weight frontier model by Q4 2026 strongly suggests a Llama 4 release. The focus on coding (Muse Spark, agentic tools) indicates coding benchmarks will be the primary competitive axis. I predict the model will surpass GPT-4o-class performance on HumanEval+ and SWE-bench by at least 5% each.
Given the AMD-Cerebras partnership's 5x T/s/W claim and NVIDIA's $500B SK Group deal for HBM4, NVIDIA cannot afford to let a competing disaggregated architecture gain traction. Expect an acquisition or deep partnership within 12 months.
The convergence of the LLM waterfall pattern (429 failover) and Offloop's D1 dispatcher (multi-agent orchestration) will coalesce into a standardized middleware layer within 2 quarters. Expect a startup or open-source project to release a unified routing protocol that abstracts provider choice, failover, and cost optimization, similar to what Kubernetes did for compute.
By October 28, 2026, Amazon will change the default model in Bedrock agent configurations from Anthropic's Claude 3.5 Sonnet to a new Amazon Nova model (likely 'Nova-Agent' or 'Nova-Pro') that is natively fine-tuned for tool-calling, MCP adherence, and multi-step reasoning. This is not a full replacement of Claude options — Anthropic models remain available — but the default shift signals Amazon's strategic move to reduce Anthropic dependency and capture more margin per inference.
Within 90 days, Anthropic will announce a second major cloud or hardware partnership beyond the AMD MI450 deal — likely with Google Cloud TPUs or Intel Gaudi — making it the only frontier lab running production inference across two distinct accelerator architectures. This will pressure OpenAI and Meta to follow suit, breaking Nvidia's single-vendor lock on frontier inference.
Google's strategy of shipping three Flash models while delaying 3.5 Pro signals a structural shift: the race is moving from frontier capability to cost-efficient capability. Within one quarter, at least two other labs (likely Meta and Mistral) will announce similar 'cheap-tier' models that match or exceed GPT-5.6 on agent benchmarks, compressing margins for inference providers and forcing differentiation into memory, orchestration (MCP), and security layers.
MCP becomes a default procurement checkbox for enterprise agent platforms. Graph evidence: 124 shared articles with Claude Code, 60 with Anthropic, and indirect links via Cloudflare and other infrastructure nodes; low bridge score today but high adjacency concentration around the dominant agent cluster.
Google will formalize a deeper agent/workflow integration around Claude Code or a Claude-adjacent developer surface. Graph evidence: High shared-neighbor count between Claude Code and Gemini; Google’s bridge score is high; Claude Code has the highest PageRank in the graph; repeated launch-response motif between Google and Anthropic.
Join readers from Google, Anthropic, and NVIDIA. Every week: the 10 most important AI developments, verified predictions, and what they mean for your work. Free forever. Customize what you get →