This week brought a cheap new Claude for everyday building, a trillion-parameter model from Europe, a whole new category of probability-based models, and useful open options for coding. For beginners, the signal is simple: budget models got better, image costs shifted, and most big models are still previews. This roundup is based on official documentation, pricing pages, and community reports — we did not run these tools hands-on — with plain verdicts to help you decide what to ignore. Start with our beginner model comparisons if you need background first.
What shipped this week
| Model/tool | Who made it | What it is | Price | Open weights? |
|---|---|---|---|---|
| Claude Haiku 5.5 | Anthropic | Fastest, cheapest Haiku for agents and chat | $0.10 input / $0.50 output per Mtok (per million tokens) up to 100k; higher above 100k | No |
| Mistral Large 4 | Mistral | 1-trillion-parameter multimodal model in preview | Launch sale $0.68 input / $2.09 output per Mtok; standard $1.36 / $4.18 | Coming by end of month |
| GLM-5.3 on Bedrock | Z.ai via AWS | Large open-weight model for enterprises on Bedrock | Enterprise Bedrock pricing varies | Yes |
| Beam | Reflection AI | Text-only open-weight coding model, waitlist only | No public price | Yes, due later in October |
| Nano Banana 2.1 | Image generation and editing update | $30 image output per Mtok; $1.50 text input per Mtok | No | |
| Mellum2.1 Thinking | JetBrains | Small open model for coding agents | Free, self-hosted | Yes |
| Clef / Clef-flash | Cloudflare | Decision models that output answers plus probabilities | $0.24 / $0.09 per Mtok input only | Yes, Apache 2.0 |
| pplx-decider-v1-27b | Perplexity | Decision model behind Decisions API | $0.04 per Mtok input only | Yes, Apache 2.0 |
Claude Haiku 5.5 makes budget AI actually good
Claude Haiku 5.5 is Anthropic’s cheapest and fastest Haiku yet, with much lower pricing and an adjustable effort setting for harder tasks. Vendor benchmarks look strong for coding agents, and an independent score puts it ahead of GPT-6 Luna. It is available through major clouds for builders on a budget.
Anthropic launched Claude Haiku 5.5 on Oct 7 as its budget option that can still handle agent work. Pricing is $0.10 input and $0.50 output per Mtok up to 100k tokens, rising to $0.50 and $2.50 above 100k, according to Anthropic pricing. That is roughly 90% cheaper than Haiku 4.5 up to 100k. Cache reads (discounted pricing when reusing stored prompts) are also very low at $0.01 and $0.05.
This is the first Haiku with an adjustable effort setting, so you can trade speed for more careful reasoning. Vendor benchmarks show 39.2% on Terminal-Bench 4.0 versus 0.0% for Haiku 4.5 and 16.4% for GPT-6 Luna, plus 72.4% on OSWorld 2.1. The independent Artificial Analysis score for Haiku 5.5 gives it 43 at max effort versus 38 for GPT-6 Luna. It is available on AWS, Google Cloud, and Azure. Anthropic also halved Sonnet 5.5 cache-read pricing and added monthly API credits for Max and Team subscribers.
Continue reading: Claude Opus 5.5 vs GPT-6 Astra vs Grok 4.7 compared
Where this fits: choose Haiku 5.5 if you pay for API use and want low cost for chatbots, summaries, or simple coding helpers.
Mistral Large 4 is Europe’s biggest model bet yet
Mistral Large 4 is a very large multimodal model in public preview on Mistral Studio, with open weights expected by the end of the month. Mistral positions it as Europe’s major frontier bet, trained in its own European datacenters. Early performance claims are vendor-reported, so treat them as preliminary.
Mistral describes Mistral Large 4 as a 1-trillion-parameter natively multimodal mixture-of-experts (MoE: many small expert sub-networks, only some active per query) with 52B active parameters. It is in public preview on Mistral Studio, with weights due by end of month. See the Mistral Large 4 model documentation for supported uses and limits.
Launch pricing is $0.68 input and $2.09 output per Mtok, moving to standard $1.36 and $4.18 after sale. Mistral says it trained on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters and supports 160-plus languages. Vendor-reported scores of 61.7% DeepSWE v1.1, 28.3% Terminal-Bench 4, and 93% Cybench are self-reported and not yet independently confirmed. The launch follows a €3B Series D, described as the largest equity round by a European tech company.
Continue reading: why small models still matter for beginners
Where this fits: wait unless you already experiment in Mistral Studio; beginners should watch for weights and independent tests.
GLM-5.3 lands on Amazon Bedrock for enterprises
GLM-5.3 is Z.ai’s large open-weight model now generally available on Amazon Bedrock for eligible enterprise customers. It offers a very long context window with OpenAI-compatible APIs, prompt caching and cross-Region inference. For beginners, this one matters mainly if your workplace already builds inside Amazon’s cloud.
AWS announced GLM-5.3 availability on Bedrock on Oct 5. It is a 753B-parameter mixture-of-experts model with 1M-token context, meaning it can take very long documents in one request. The Bedrock model card for GLM-5.3 lists OpenAI-compatible APIs, prompt caching, and cross-Region inference for eligible enterprise customers.
There is no simple consumer signup here. Access runs through Bedrock enterprise controls, which handle permissions, logging, and regional routing. That makes it more relevant to teams than to solo beginners.
Continue reading: Qwen Code CLI tool review for coding beginners
Where this fits: choose this path only if your team already uses Bedrock and needs long-context enterprise deployment.
Reflection’s Beam is an open-weight coding bet, not a product
Reflection AI’s Beam is a large text-only open-weight model focused on coding, currently waitlist-only with no public API or price. The company says its coding performance sits near GLM-5.2 and behind newer rivals, which is self-reported. Weights and a technical report are promised later in October.
According to Reflection’s Beam announcement, Beam is a 501B total and 23B active text-only mixture-of-experts model. There is no public API or price yet, and access is waitlist only. Planned Apache 2.0 weights and a technical report are due later in October.
Reflection says Beam is roughly level with GLM-5.2 and behind GLM-5.3, Kimi K3, and Qwen3.8 Max on coding. Those comparisons are self-reported, so they should not guide a purchase decision. For now, Beam is a research preview to watch, not a tool to adopt.
Continue reading: beginner’s guide to running local models with Ollama
Where this fits: skip for now unless you track open coding models; wait for public weights and independent scores.
Nano Banana 2.1 halves image prices, quietly
Nano Banana 2.1 is Google’s updated image generation and editing model, with much cheaper image output but higher text input pricing. The older Nano Banana 2 is deprecated, though Google has not announced a shutdown date. If you generate images with Gemini, check your cost mix before switching.
Per the Gemini API changelog, gemini-nano-banana-2.1 is the new image generation and editing version, and Nano Banana 2 is deprecated with no announced shutdown date. Do not assume a retirement date — none was announced.
Pricing details in the Gemini API pricing page show image output halved to $30 per million tokens, or $0.0336 per 1K image versus $0.067 before. At the same time, text input rose from $0.50 to $1.50 per Mtok. If you mostly output images, costs drop. If your prompts include lots of text or reference images, the input increase offsets some savings.
Continue reading: Gemini 3 Flash review for everyday use
Where this fits: pick 2.1 for new image projects, but review input-heavy editing workflows where higher input pricing applies.
JetBrains Mellum2.1 Thinking targets coding agents
JetBrains Mellum2.1 Thinking is a small, fast open model built specifically for coding agents, with weights and local-ready files published on Hugging Face. It is designed to handle routine agent steps quickly and cheaply. For beginners curious about local coding help, this is the easiest low-risk entry this week.
JetBrains announced Mellum2.1 Thinking for coding agents on Oct 8 as a 12B total and 2.5B active model for agent steps like edits and tool calls. Weights and the Mellum2.1 collection on Hugging Face include GGUF (a file format for running models locally), which helps for local testing.
Because it is small and open, it can run with lower memory and cost than frontier models. It will not match large models on hard reasoning, but that is not the point — it is for fast, repeatable coding assistance inside agent loops.
Continue reading: Ollama vs LM Studio for local models
Where this fits: try it if you want a free local coding helper; for broader advice on code checks, see this guide to practical code quality checks.
Decision models: AI that answers with a probability, not prose
Decision models are a new category that return a structured answer plus a probability instead of chatty text, and bill only for input. This week Cloudflare launched Clef models, Perplexity launched a decider model, and AWS published a small self-hosted recipe. They suit classification and routing, not everyday chatting.
Cloudflare shipped Clef (27B) and Clef-flash (9B) under Apache 2.0 as its first in-house-trained models, running on Workers AI at $0.24 and $0.09 per Mtok with output free, per the Cloudflare Clef announcement. Perplexity shipped pplx-decider-v1-27b at $0.04 per Mtok behind its Decisions API, documented in the Perplexity Decisions quickstart, with weights at pplx-decider on Hugging Face. AWS Strands Labs added Strands Decider 2B with a self-hosted training recipe, described in Introducing Strands Decider.
Think yes-or-no sorting, spam triage, or routing a ticket — not essays or brainstorming.
Explore more: compare live models in the arena
Where this fits: use a decision model if you need cheap classification at scale; for workflow ideas, see this overview of no-code automation patterns.
GitHub trending: agent memory and reach tools
Two open-source agent tools trended on GitHub this week for practical reasons. One adds persistent memory across coding sessions, while the other lets agents search social and video platforms through one command-line tool. Both are community projects for tinkerers, not polished products for complete beginners.
The repo thedotmack/claude-mem adds persistent memory across agent sessions and works with Claude Code, Codex, Gemini, and Copilot. The repo Panniantong/Agent-Reach lets agents read and search X, Reddit, YouTube, GitHub, and Bilibili via one CLI.
Both are early-stage and require comfort with terminals, keys, and frequent updates. Memory helps agents recall project context, while reach helps gather outside information without juggling five integrations.
Continue reading: past GitHub trending repos for agents and Claude Code review for beginners
Where this fits: experiment only if you already use coding agents; otherwise watch, and see this agent development field guide for context.
Week ahead
Watch for Mistral Large 4 open weights by end of month, Beam’s Apache 2.0 weights and technical report later in October, and independent benchmarks catching up to Haiku 5.5. If those land, beginner picks for cheap coding and local use should get clearer.
Frequently asked questions
Which of this week’s models should a beginner actually use?
If you pay for API access, start with Claude Haiku 5.5 for low-cost chat and coding help through a supported provider. If you want free and local, try JetBrains Mellum2.1 Thinking via Hugging Face files. Skip waitlists and enterprise-only previews until weights and independent results arrive.
What is a decision model and do I need one?
A decision model returns a typed answer plus a probability, such as approved or spam with a confidence score, instead of free text. Billing is input-only because output is minimal. You need one only for classification, filtering, or routing at scale — not for normal chat, writing, or coding.
Are Mistral Large 4 and Beam free to use yet?
No. Mistral Large 4 is in paid public preview on Mistral Studio, with open weights expected by end of month. Beam is waitlist-only with no public API or price, and its Apache 2.0 weights are due later in October. Treat both as items to watch, not tools to adopt today.
📖 Related Reads
- ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
- CodeIntel Log — code quality, debugging, and software engineering benchmarks
- NoCode Insider — AI workflow automation with no-code tools, agents, and APIs
- Hermes Tutorials — Hermes Agent setup, configuration, and advanced workflows
Cross-links automatically generated from None.
Back to all posts