A Weekend Earthquake for Coding Agents: DeepSeek Price Cuts and Cursor Composer taking on Opus

A lot can change in 48 hours. Two days after writing about Qwen3.7 Max, Gemini 3.5 Flash, and Antigravity limits, the weekend turned into a mini earthquake for coding agents: Anthropic roadmap leaked further, DeepSeek effectively repriced frontier class compute, Cursor Composer 2.5 showed up swinging at Opus 4.7, and Hermes Agent quietly became a gateway to good enough frontier for almost no money.

Below is a walkthrough of what shifted and what it means if you live inside agent harnesses all day.

Anthropic 4.8 Leak: Roadmap, Not Release

The Opus 4.8 rumor floating around is really a Sonnet story: the latest digging into the Claude Code source map leak shows explicit references to a Sonnet 4.8 string, plus internal codenames like Capybara, Mythos, and a cluster of hidden feature flags for advanced agent behavior.

The important part: so far there is still no official Anthropic announcement, no model card, no API ID, and no pricing for any Sonnet 4.8 model. Markets that tried to bet on a May 2026 Sonnet 4.8 launch priced that outcome near zero by late May, which is a decent proxy for the leak being real, but timing being fantasy.

The more sober read from folks who went through the leak is clear. Opus 4.7 is real and shipping; Sonnet references in the codebase are planning markers, not release guarantees. The leak also surfaced flags for things like always on autonomous modes and advisor patterns where a stronger model shadows a weaker one for oversight.

For harness builders, the leak matters less as Opus 4.8 tomorrow and more as confirmation that Anthropic is architecting deeper agent features into Claude Code itself, but nothing you can target in your stack yet.

DeepSeek V4 Pro: The 75 Percent Discount Becomes the Price

DeepSeek took its already aggressive V4 Pro promotion and made it the new normal. What started as a 75 percent time limited discount is now announced as the post promo baseline for the API.

The before and after looks roughly like this:

Input cache miss: $1.74 to $0.435 per 1M tokens

Input cache hit: $0.145 to $0.0036 per 1M tokens

Output: $3.48 to $0.87 dollars per 1M tokens

Earlier analysis already pegged DeepSeek V4 Pro as the cheapest of the big models, undercutting Gemini 3 series, GPT 5.4, and Opus even before this permanent cut.

DeepSeek extended the 75 percent promo to the end of May and then stated that once the window closes, the discounted numbers simply become standard pricing, not a limited sale. For agent workloads, that effectively locks in a flagship class model at about one quarter of its original cost, and already cheaper than most mini frontiers on a pure token basis.

If you are paying for long running autonomous loops, this is the kind of move that changes spreadsheets: V4 Pro now sits in the “I can leave this agent on overnight without panic” bracket.

Cursor Composer 2.5 positioning against frontier models on the Intelligence vs. Cost graph (Artificial Analysis)

Detailed table comparing Composer 2.5 against Opus 4.7 and GPT-5.5 on agent benchmarks and costs

Composer 2.5: Cursor House Model Punches at Opus 4.7

Cursor pushed Composer 2.5 and, for the first time, their in house model is squarely in Opus 4.7 territory on agent benchmarks while costing a tiny fraction per task.

Artificial Analysis now ranks Composer 2.5 at 62 on the Coding Agent Index, a 14 point jump over Composer 2, landing it just behind heavier effort Claude Opus 4.7 and GPT 5.5 variants in the same index band. On detailed benchmarks: Terminal Bench style scores: Composer 2.5 Fast comes in within a rounding error of Opus 4.7. Multi benchmark rollups show Composer 2.5 trailing Opus 4.7 by only a few points across categories.

The killer part is the economics per task. Composer 2.5 standard in Cursor hits an Index of 62 for just 0.07 dollars typical cost per task, with token pricing at 0.50 and 2.50 dollars. Composer 2.5 Fast maintains that 62 Index for 0.44 dollars typical cost per task. Compare this to Claude Opus 4.7 which scores a slightly higher Index of 66 but costs a massive 4.10 dollars per task, or GPT 5.5 which scores 65 and costs 4.82 dollars per task.

Artificial Analysis points out that Composer 2.5 is now the cheapest agent scoring above 60 on their Coding Agent Index, with medium effort peers costing 10 to 60 times more per task depending on the variant.

In practical harness terms: if you are comfortable living inside Cursor and its ecosystem, Composer 2.5 just became a default workhorse that can handle Opus level debugging and code review without Opus level bills.

Hermes Agent and DeepSeek: Frontier Enough Agents for Almost Free

On the orchestration side, Hermes Agent continues to morph into the go to self improving harness, and the DeepSeek story around it is getting wild.

A few converging threads: DeepSeek V4 Flash is exposed as a free model on multiple fronts, including a dedicated free tier on OpenRouter with a 1M token context and roughly 49.2B free tokens per week. DeepSeek V4 Flash and Pro are also runnable locally via Ollama and Hugging Face, letting you spin up Hermes style agents with zero per token API cost if you are willing to manage your own GPU or beefy CPU box. Nous ecosystem has a promo where DeepSeek V4 Flash is completely free inside the Nous Portal for a limited time, explicitly pitched as god tier for no money. Hermes provider docs show straight through integration with DeepSeek own API, OpenRouter, and self hosted endpoints.

Combine that with DeepSeek permanent V4 Pro price cut and the free Flash pathways, and you get a weird new equilibrium: you can run a self improving, tool rich agent harness on near frontier models for effectively free in the margin. This encroaches directly on use cases that, even a few months ago, felt like the exclusive territory of closed IDE copilots.

GPT 5.6: The Next Ghost on the Horizon

Meanwhile, OpenAI next flagship keeps hovering as a moving target. Recent coverage talks about GPT 5.6 candidates going through rapid internal testing, lumped together with Gemini 3.2 Pro leaks and Mythos V2 experiments.

Separate rumor summaries peg the likely GPT 5.6 launch window as late 2025 into early 2026, based on infrastructure spend and hiring patterns, but make it clear that dates are speculative and not backed by any OpenAI announcement.

For practical harness decisions today, GPT 5.6 is more pressure on everyone else roadmap than something you can plug into. The implication is that the cost per reasoning token frontier keeps marching forward, but the models we actually have in our hands this weekend are Qwen3.7 Max, GPT 5.5, Opus 4.7, Gemini 3.5 Flash, and now DeepSeek V4 and Composer 2.5 pushing hard from below.

The Weekend Big Pattern: Agents Got Cheaper Again

Put all of this next to the last article changes, and a pattern emerges: every lab is undercutting the effective price of good enough agent brains, either via discounts, house models, or free tier funnels.

If you are optimizing for agentic coding rather than pure chat UX, a reasonable stack after this weekend might look like: Composer 2.5 inside Cursor as the default interactive coding brain when you are in live co dev mode, where cost per task and latency matter more than brand names. DeepSeek V4 Pro as the bulk agent worker for long running jobs, CI style automation, and batch refactors, now that its permanent pricing makes overnight runs financially boring. DeepSeek V4 Flash and Hermes Agent for self improving, tool heavy agents that you want to run either truly free or at negligible cost, especially for glue tasks and background workflows. Opus 4.7 and GPT 5.5 reserved for the narrow band of problems where their last few percentage points on hard benchmarks actually translate into real value, rather than as default hammers.

The Anthropic leak and GPT 5.6 rumors are mostly a reminder that the ceiling is still moving; the DeepSeek and Cursor moves are what actually change this week invoices.

In other words: two days after I wrote that Qwen3.7 Max and Gemini 3.5 Flash finally gave us four top shelf coding brains, the weekend quietly added at least two more serious contenders on the agent per dollar frontier, and one of them is flirting with effectively free if you are willing to wire it into the right harness.