ChatGPT Goes Interactive as Haiku Cuts Prices and Google Bets on Work Agents

Thursday, 8 October 2026 | Coverage window Since the 7 October edition (18:46 IST) | Compiled 18:46 IST

OpenAI widened its interactive ChatGPT rollout, Anthropic lowered small-model prices and Google announced a persistent workplace agent. New coding, retrieval and containment tools arrived as a disputed teen-safety assessment put deployed safeguards under scrutiny.

1. GPT-6 reaches more ChatGPT users with interactive Intelligent UI

REPORTING OpenAI began rolling out GPT-6 with Intelligent UI on 7 October to Plus, Pro, Business and Enterprise, with Free and Go following on 8 October. Replies can contain charts, forms, buttons and working tools rather than only prose. Paid Chat tiers use GPT-6 Sol; Free and Go use Luna. Enterprise access depends on admin settings. This changes Chat, not the models behind Work or Codex.

Analysis A wider distribution and interface change, not a new Astra release. Generated interfaces could make small tasks easier, but layout quality and factual accuracy still need checking. OpenAI's speed and safety improvements are vendor claims.

OpenAI / TechTimes · 7-8 Oct 2026 · Primary rollout details · Independent coverage

2. Claude Haiku 5.5 lowers small-model prices and adds effort controls

REPORTING Anthropic released Haiku 5.5 with adjustable effort and a 1M-token context window. Prompts up to 100K cost $0.10 input and $0.50 output per million tokens; above that threshold, both rates rise fivefold. Anthropic also halved Sonnet 5.5 cache-read prices. Artificial Analysis independently scores Haiku at 43 on its Intelligence Index at max effort, but observes roughly three times Luna's output-token use per task.

Analysis A strong option for bounded subagent work, summaries and high-volume tasks. Cheap token prices do not guarantee cheap tasks: effort, token use and the long-prompt price step matter. Vendor claims of average savings and benchmark wins are not universal workload results.

Anthropic / Artificial Analysis · 7 Oct 2026 · Primary release and pricing · Independent evaluation

3. Google announces a persistent Gemini agent for enterprise work

REPORTING At Gemini at Work, Google introduced a universal agent that plans work, uses tools and business context, and chooses models with cost controls. VentureBeat describes long-running workflows, Claude support and a coworker configuration with its own Workspace identity, email, calendar and Drive. Google promises attested identities, audit logs, policy enforcement and sandboxed execution. Activation, licensing and complete availability details remain unspecified.

Analysis Google is competing to become the central delegation surface for workplace tasks. A dedicated account is one configuration, not required for every use. Treat the launch as an announcement of capabilities, not proof every customer can deploy all of them today.

Google / VentureBeat · 8 Oct 2026 · Primary announcement · Independent capability and availability review

4. JetBrains releases Mellum2.1 for self-hosted coding agents

REPORTING JetBrains released Mellum2.1, a 12B mixture-of-experts model with 2.5B active parameters under Apache 2.0. Reinforcement learning in real environments targets repository exploration, file edits and checking changes. The architecture is unchanged from Mellum2; most work went into post-training. Weights are on Hugging Face. GGUF builds and the multi-token-prediction head for vLLM are described as coming soon.

Analysis An open, compact worker model for local coding and subagent systems. JetBrains' benchmark and throughput comparisons use its own setup; claimed speed on an H200 does not establish laptop performance. The promised deployment packages are not yet the release.

JetBrains · 8 Oct 2026 · Primary release

5. Windows makes agent containers generally available and plans local-cloud coding

REPORTING Microsoft made Execution Containers generally available on Windows 11, with runtime policies for file and network access and support from several coding agents. It also announced local MAI Code 1.1 Flash, a 137B-total/6.8B-active model, and plans for HydraFusion to route between local and cloud inference in Copilot, CLI and VS Code. That Windows hybrid-routing experimental preview is due later in October.

Analysis Containment and identity are becoming operating-system features rather than add-ons. Separate the shipped container capability from planned routing, hardware and Copilot features. HydraFusion already existed; local Windows routing is the new extension, not a first launch of the orchestrator.

Microsoft / Windows developer blog · 7 Oct 2026 · Primary Windows announcement · Coding and sandbox details

6. Perplexity releases interoperable late-interaction embedding models

REPORTING Perplexity released pplx-embed-v2-late in 0.6B and 9B sizes for text, images and rendered document pages. Both share a token-level embedding space, so the small model can query a large-model index. MaxSim compares 128-dimensional vectors per token rather than one vector per document. Weights are available on Hugging Face under MIT. Perplexity reports 92.4% for the 9B model on MADQA; a hosted API is planned, not live.

Analysis Useful for visual-document RAG and lower-cost query encoders without rebuilding a quality-focused index. More token vectors mean more storage and scoring work. Scores are Perplexity's own, and superiority on one benchmark is not a general retrieval verdict.

Perplexity / MarkTechPost · 7 Oct 2026 · Primary technical release · Deployment and benchmark review

7. SkillForge trains agents to retire obsolete skills, not only collect them

REPORTING A new paper proposes co-evolving agent policies and skill libraries through trial, active, stable and retired states. Base-model rollouts pre-retire weak skills before supervised fine-tuning; reinforcement learning continues retirement and mutation. Authors report up to 7.8% relative improvement over their strongest baseline and release SkillFurnace, a dataset of more than 5,000 annotated records. The paper appeared on 7 October and was submitted by its author to Hugging Face on 8 October.

Analysis The research targets a real failure mode: once-useful procedural memory can become harmful as a model improves. Results are authors' experiments, not independent reproduction. This is the October co-evolution paper, distinct from an August paper with the same SkillForge name.

arXiv / Hugging Face · 7-8 Oct 2026 · Primary paper · Publication and author submission

8. Manus raises more than $500M after resuming independence from Meta

REPORTING CNBC reports that Manus parent Butterfly Effect completed a round exceeding $500 million, led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund participating. It is the first round since the Meta acquisition was unwound following Chinese regulatory intervention. The company did not disclose its post-funding valuation; the $4 billion figure comes from Bloomberg's earlier reporting, not confirmation of this round's terms.

Analysis Capital is still backing independent agent products even as model prices fall. The more consequential question is durable distribution and execution economics. Do not present the earlier reported valuation or regulatory history as new deal terms.

CNBC / DealStreetAsia · 8 Oct 2026 · Independent company-statement coverage · Funding corroboration

9. Common Sense Media challenges ChatGPT's teen safeguards; OpenAI disputes testing

REPORTING Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an 'Unacceptable Risk' after testing over 4,000 prompts. It reports failed parental alerts on new linked accounts, inconsistent crisis referrals and easy study-mode bypasses, while some protections worked. AP reports OpenAI's response: much testing may have ended before parental controls finished activation. The institute discloses industry funding, including the OpenAI Foundation, and asserts editorial independence.

Analysis A material safety dispute about deployed protections, not a settled regulatory finding. Independent testing is valuable, but activation state and methodology must be reconciled before treating these results as representative of every teen account or today's model rollout.

Common Sense Media / AP · 7 Oct 2026 · Primary assessment announcement · Independent coverage and OpenAI response

10. Rein Security raises $25M for runtime protection of AI agents

REPORTING SecurityWeek reports Rein Security raised $25 million in Series A funding, bringing total funding to $35 million. Glilot Capital and Sienna Venture Capital co-led. Rein extends runtime application security to agents, offering behavior visibility, real-time guardrails, governance and supply-chain protection. The company says it secures thousands of agents; funding will support product development, research and hiring.

Analysis Agent security is attracting investment around execution rather than just prompt filtering. Deployment counts and protective effectiveness are company claims, not independent audits. Funding demonstrates investor demand, not that the platform blocks every harmful action.

SecurityWeek · 8 Oct 2026 · Independent funding coverage