New releases are widening the range between frontier capability and low-cost throughput. Claude Sonnet 5's tokenizer complicates its introductory price, Reuters finds agent loops inflating completed-task bills, and Google's low-cost image model trades faster iteration for visible quality limits.
1. Claude Sonnet 5 targets agent workloads at a lower price
Anthropic released Claude Sonnet 5 across Claude plans, Claude Code, and its API. Introductory API pricing is $2 per million input tokens and $10 per million output tokens through August 31; standard pricing then rises to $3 and $15, respectively.
Anthropic's benchmark results show gains over Sonnet 4.6 in reasoning, tool use, and coding, but those remain vendor measurements. The updated tokenizer can turn the same content into roughly 1.0 to 1.35 times as many tokens, so completed-task cost may move differently from the advertised token rate.
Sources: Anthropic's Claude Sonnet 5 announcement
2. Falling token prices are not containing total AI bills
AI unit prices keep declining, yet longer contexts and multi-step agents are making completed tasks more expensive and less predictable. Reuters reported that open-source traffic on OpenRouter rose from 34% in January to 65% in June as businesses looked for cheaper alternatives and more flexible routing.
The practical metric is cost per accepted outcome, including retries, tool calls, and human review. Separate instrumentation for those components enables routine work to reach lower-cost models while reserving premium inference for steps where internal evaluations demonstrate a material quality gain.
Sources: Reuters on enterprise AI cost pressure
3. Nano Banana 2 Lite prioritizes fast, inexpensive image iteration
Google released Gemini 3.1 Flash-Lite Image, branded Nano Banana 2 Lite, for high-volume generation and editing. Google lists four-second text-to-image latency and a price of $0.034 per 1K-resolution image, positioning it as the replacement for Gemini 2.5 Flash Image.
The lower-cost model retains editing and character-consistency features, but Google notes weaknesses in small faces, spelling, fine detail, complex edits, and factual graphics. It fits prototyping and bulk variants; customer-facing publication remains gated by visual review, provenance controls, and fact checking.
Sources: Google's developer release · Google DeepMind model page