http://softwareplanetgroup.co.uk/

Software Planet Group, Austen House, Station View, Guildford (2026)

08/10/2026

The IT industry is changing at a breathtaking pace. What worked yesterday may already be outdated today, and predicting where technology is heading has never been more challenging.

One of the best ways to keep your finger on the pulse is to stay connected with the tech community - meet the people driving change, exchange ideas, and see what the industry is actually working on.

We've put together a selection of upcoming tech expos and conferences taking place in London and across the UK in late 2026 and early 2027. If you want to stay ahead of the curve, these are the events worth keeping an eye on.

Check out our selection: 👇
https://softwareplanetgroup.co.uk/tech-events-you-wont-want-to-miss-in-2026-2027/

22/09/2026

Grok 4.7: Near-Opus Coding at a Fraction of the Price

Every few weeks a new frontier model arrives claiming to close the gap with the leaders. Most of the time the gap turns out to be wider than the launch post suggests. Grok 4.7 is more interesting than usual, and engineering leaders should take an afternoon to evaluate it. What they should not do is rewrite their AI strategy around a benchmark table.

What shipped

xAI, which now brands itself SpaceXAI on official pages, released Grok 4.7 on 21 September 2026, with day-one availability in Cursor, Grok Build, the Grok API and third-party harnesses, routers and clouds. It has a 500,000-token context window, accepts text and image input.

The price is the headline

Grok 4.7 costs $2.00 per million input tokens and $6.00 per million output, against $5.00 and $25.00 for Claude Opus 5. That is 2.5 times cheaper on input and more than four times cheaper on output. For agentic workloads, which generate a lot of output, this difference matters.

On Cursor's coding benchmark the two models are nearly level. On CursorBench 4.0, Grok 4.7 at xhigh scored 46.3% at an average of $6.01 per task.

The caveats that matter

The broader picture is less flattering. On the Artificial Analysis index, Grok 4.7 trails Opus 5 by 46 to 51, and it used 1.7 times as many output tokens across the evaluation. Artificial Analysis flagged it as unusually verbose. A cheaper token is not a cheaper task if the model needs far more of them to reach a worse answer.

Long context changes the maths too. Grok 4.7's pricing steps up to $4.00 and $12.00 at or above 200k prompt tokens, whereas Opus 5 bills its full 1M window at flat rates. For large-codebase analysis or long document work, the advantage narrows sharply.

Summary

Grok 4.7 is a credible, cost-efficient option for coding workloads and deserves a place in your evaluation set. As ever, the teams that benefit most will be the ones that can test a new model against their own work within a day, not the ones reacting to launch-day headlines.

17/09/2026

Everyone knows the basic rule of optimisation: before improving anything — even a single SaaS feature — you first profile it, find the real bottlenecks, and only then decide where to invest effort.

The same principle works remarkably well with AI-assisted development. Use ChatGPT first to analyse the project: give it logs, usage data, event history, code context and other read-only sources. Ask it to separate confirmed problems from assumptions, identify bottlenecks, compare possible approaches and prepare a concrete implementation plan.

Then hand that plan to Codex for ex*****on. In other words: ChatGPT for profiling, reasoning and deciding what should change. Codex for implementing the chosen change. Without the first step, AI development can easily become optimisation by intuition. And as with software performance, intuition is rarely a substitute for profiling.

06/09/2026

Today we used OpenAI’s Batch API for a client project for the first time.

We’d known about the feature for a while, but this was our first real production use case - and it’s a very useful option when cost matters more than immediate response time.

The idea is simple: instead of sending API requests synchronously and waiting for each response, you submit a batch of requests for asynchronous processing. OpenAI has a completion window of up to 24 hours, and the cost is 50% lower than using the equivalent synchronous API.

You don’t necessarily have to wait 24 hours - the batch may complete much sooner - but you design the workflow assuming that the results are not required immediately.

This makes a lot of sense for workloads such as processing, extracting or classifying large volumes of data, where you may have thousands or millions of items to handle and no business reason to get every answer in real time.

If you’re processing a substantial amount of data with LLMs and don’t have an unlimited API budget, the Batch API is definitely worth considering.

Sometimes the easiest optimisation isn’t choosing a smaller model. It’s simply deciding that not everything needs an answer right now.

02/09/2026

Anthropic has switched on invisible watermarking for all text generated by Claude - globally, not just in the EU where the law demands it. Other major providers have signed the same code of practice, so within a year most AI-generated text online will carry a maker's mark that survives copy-paste, conversion and even retyping.

If your company uses AI in proposals, documentation, marketing or code, this deserves 10 minutes of your attention - not because there's anything to hide, but because misconceptions are already spreading. What does a watermark actually reveal? Which workflows leave a strong trace and which leave almost none? And why can no detector - including this one - prove that a human didn't do the work? We've broken down the mechanics, without the hype, here:

https://softwareplanetgroup.co.uk/ai-watermarking-is-here/

01/09/2026

Choosing an LLM is no longer a one-time technology decision.

The market is moving far too quickly for that. New models appear, existing ones improve, prices change, and older models are eventually deprecated. Something that worked perfectly well six months ago may suddenly stop working, become too expensive, or simply start producing results that are no longer competitive.

So there are two practical problems to solve. First, switching models should be inexpensive. Your application should not be tightly coupled to a single provider or model. Services such as OpenRouter (openrouter.ai) can provide a useful abstraction layer, allowing teams to experiment with and migrate between different models without repeatedly rebuilding the integration layer. Second, model selection needs to be continuous. Not every company has the budget to integrate and extensively test every new GPT, Claude, Gemini, DeepSeek or other model that appears. More importantly, this testing should not happen only when the original architecture is designed. It should continue throughout the lifetime of the product.

Specialised platforms such as Arena (arena.ai) are useful here. They provide an independent way to compare models and understand how different LLMs perform across different types of tasks.

There is rarely a universally “best” LLM. There is a model that is currently the best fit for your particular combination of quality, reasoning, latency, context size and cost. And that answer can change surprisingly quickly.

29/08/2026

12 November 2026 could be the day Cursor loses its built-in GPT models.

OpenAI is terminating its agreement with Cursor following the company's acquisition by SpaceX. And this is probably one of the most interesting break-ups we have seen in the AI industry so far. Formally, everything is very civilised. OpenAI has proposed the maximum transition period allowed under the agreement, apparently to give developers time to adapt. But the explanation behind the decision is considerably less polite.

OpenAI says it cannot be confident that its technology, when used by Elon Musk-controlled companies, will continue to be used in accordance with its terms. And then comes the history. After Musk acquired Twitter, OpenAI says contractual terms were violated. OpenAI also points to Musk's testimony that xAI used OpenAI outputs in the development of its own models.

In other words, the neighbour has already been caught looking inside your fridge, and is now surprised that you have changed the lock. There is another particularly interesting detail. In the same announcement, OpenAI mentions the name of an upcoming model: Astra. It also makes it clear that future models will not be made available to Cursor.

And the proposed Cursor cut-off date is 12 November. Pure speculation on our part, but the coincidence is hard to ignore: could 12 November also be connected with Astra's launch?

If so, the message would be rather brutal. Cursor keeps the existing models during the transition period, but gets thrown overboard just as the next generation arrives.

There is obviously a personal history behind all of this as well. Sam Altman and Elon Musk once founded OpenAI together. Today, their disagreement is being played out through lawsuits, contracts, competing AI companies and now access to models. But the people who ultimately feel the consequences are developers.

If a large part of your engineering workflow depends on Cursor and GPT models, a disagreement between two billionaires can suddenly become your infrastructure problem.

Fortunately, this is not an emergency. Cursor users can still use their own OpenAI API keys, and there are other models available inside the editor. But there is a broader engineering lesson here. AI models are becoming infrastructure. And just like databases, cloud providers or external APIs, they create dependencies.

At Software Planet Group, we think AI-assisted development processes should be designed with this in mind: models should be replaceable components, not foundations that the entire engineering workflow depends on.

Because apparently, even your IDE can now become collateral damage in someone else's divorce.

19/08/2026

We certainly never expected the day when the Software Planet Group blog would be giving clients advice on how to build software themselves with generative AI. Yet here we are.

The competitive landscape has changed. We are no longer competing only with other bespoke software development companies. In some cases, we are also competing with our own clients, who can now turn an idea into working code or an MVP prototype before they ever speak to an engineering team. And we actually see that as a positive development.

Increasingly, clients come to us with much more than an idea. They bring prototypes, AI-generated codebases and early versions of real products. That means our role can move further towards the work where experienced engineering judgement matters most: understanding what has been built, identifying architectural and quality risks, preserving what is valuable, and turning an experimental product into software that can be maintained, scaled and developed with confidence.

Many development companies would rather start again from a clean repository. We have always been comfortable doing something harder: working with existing code, stabilising it, refactoring it and gradually turning it into a strong product without unnecessarily throwing away what already works.

Perhaps generative AI does not remove the need for software engineering. It simply moves the starting line.

The article turned out to be rather more extensive than we initially expected, so we’ll be publishing it in several parts. Keep an eye out for the next instalment.

https://softwareplanetgroup.co.uk/planning-software-when-ai-writes-it/

08/07/2026

The pace of change in agentic AI is unlike anything we've seen since the early days of Agile. Six months ago, "harness" was a niche term; today it's how the most effective engineering teams run AI agents in production — and at a recent hackathon, 19 of the top 20 teams had built their solutions on one. If you blinked, you may well have missed the shift.

And that's the problem. For many engineering leaders, the landscape has moved so quickly — from prompt engineering to context engineering to harness building in barely two years — that it's genuinely hard to tell where the industry is heading, and which of today's approaches will still matter next quarter.

Our latest article is an attempt to bring some clarity. We trace how agents used in software development have conceptually evolved: from simple chat models, through the elaborate scaffolding era of LangChain and LangGraph, to the surprising simplification that produced modern harnesses like Claude Code and Hermes. We explain why skills — not bigger models or longer context windows — are the evolutionary point the industry has just reached, what makes them fundamentally different from the MCP-style tooling of a year ago, and where the road leads from here (hint: it involves agents that maintain and improve their own capabilities).

If you're deciding how AI agents should fit into your delivery process — or simply want to understand what your engineers mean when they say "harness" — this one's for you.

https://softwareplanetgroup.co.uk/ai-agent-skills-new-memory/

30/06/2026

For years, we viewed AI coding tools with healthy scepticism. They were impressive demonstrations, useful in places, but nowhere near capable of changing how professional software engineering teams actually worked. We kept asking ourselves the same question: Is this real yet?

At some point over the past year, without any announcement or grand revelation, our answer changed. AI stopped being just another productivity tool and became capable of delivering substantial end-to-end engineering work. We believe we have quietly crossed a threshold where autonomous engineering tasks are becoming self-contained building blocks from which increasingly sophisticated software systems can be assembled. The implications extend far beyond software development itself.

In our latest article, we explain why we believe this is not simply another technological improvement, but the beginning of a fundamental shift in what software engineering is becoming.

https://softwareplanetgroup.co.uk/the-end-of-yesterdays-software-development/

Software Planet Group

08/10/2026

22/09/2026

17/09/2026

06/09/2026

02/09/2026

01/09/2026

29/08/2026

19/08/2026

08/07/2026

30/06/2026

Address

Telephone

Website

Alerts

Contact The Business

Shortcuts

Share

Category

Listing takedown request